A Study of an Iterative User-Specific Human Activity Classification Approach

(1)

,

STOCKHOLM SWEDEN 2019

A Study of an Iterative User- Specific Human Activity

Classification Approach

NIKLAS FÜRDERER

KTH ROYAL INSTITUTE OF TECHNOLOGY

SCHOOL OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE

(2)

School of Electrical Engineering and Computer Science Kungliga Tekniska Högskolan

A Study of an Iterative User-Specific Human Activity Classification Approach

M

ASTER

T

HESIS

EIT ICT Innovation and Data Science

Author: Niklas Fürderer

Academic Supervisor: Henrik Boström

Industrial Supervisors: Arnrun Thorsteinsdottir & Tor Carlsson Examiner: Mira Kajko-Mattsson

Course 2017-2018

(3)

(4)

Acknowledgements

Firstly, I would like to express my sincere gratitude to my examiner Associate Professor Mira Kajko-Mattsson and my supervisor Professor Henrik Boström for their continuous assistance, immense knowledge, and beneﬁcial feedback throughout the whole course. Their guidance and encouraging reviews helped me particularly in conducting my research and improving my scientiﬁc writing.

I thank Noomi AB (formerly Aiﬂoo AB) and all accompanying employees and colleagues for the challenging chance to be part of this unique journey and all given and upcoming opportunities.

I would like to give special thanks to my industrial supervisors Arnrun and Tor who welcomed and integrated me in the best way possible and who supported me with their constructive and thought-provoking impulses during the entire planning and development phase. This work could not have been realized without their insightful comments and extensive encouragement.

Last but not least, I would like to thank my friends and fellow students - who became the very same during the last two unforgettable years - for their valuable inputs, their constantly honest feedback, and our inspiring fundamentally shared zest of life.

(5)

(6)

v

Abstract

Applications for sensor-based human activity recognition use the latest algorithms for the detection and classification of human everyday activities, both for online and offline use cases. The insights generated by those algorithms can in a next step be used within a wide broad of applications such as safety, fitness tracking, localization, personalized health advice and improved child and elderly care.

In order for an algorithm to be performant, a significant amount of annotated data from a specific target audience is required. However, a satisfying data collection process is cost and labor intensive. This also may be unfeasible for specific target groups as aging effects motion patterns and behaviors. One main challenge in this application area lies in the ability to identify relevant changes over time while being able to reuse previously annotated user data. The accurate detection of those user-specific patterns and movement behaviors therefore requires individual and adaptive classification models for human activities.

The goal of this degree work is to compare several supervised classifier performances when trained and tested on a newly iterative user-specific human activity classification approach as described in this report. A qualitative and quantitative data collection process was applied. The tree-based classification algorithms De- cision Tree, Random Forest as well as XGBoost were tested on custom based datasets divided into three groups. The datasets contained labeled motion data of 21 volunteers from wrist worn sensors.

Computed across all datasets, the average performance measured in recall in- creased by 5.2% (using a simulated leave-one-subject-out cross evaluation) for algorithms trained via the described approach compared to a random non-iterative approach.

Key words: human activity recognition, classiﬁcation, random forest, xgboost, decision tree, iterative learning approach, user-speciﬁc

(7)

Sammanfattning

Sensorbaserad aktivitetsigenkänning använder sig av det senaste algoritmerna för detektion och klassiﬁcering av mänskliga vardagliga aktiviteter, både i upp- och frånkopplat läge. De insikter som genereras av algoritmerna kan i ett nästa steg användas inom en mängd nya applikationer inom områden så som sä- kerhet, träningmonitorering, platsangivelser, personiﬁerade hälsoråd samt inom barn- och äldreomsorgen.

För att en algoritm skall uppnå hög prestanda krävs en inte obetydlig mängd annoterad data, som med fördel härrör från den avsedda målgruppen. Dock är datainsamlingsprocessen kostnads- och arbetsintensiv. Den kan dessutom även vara orimlig att genomföra för vissa speciﬁka målgrupper, då åldrandet påver- kar rörelsemönster och beteenden. En av de största utmaningarna inom detta område är att hitta de relevanta förändringar som sker över tid, samtidigt som man vill återanvända tidigare annoterad data. För att kunna skapa en korrekt bild av det individuella rörelsemönstret behövs därför individuella och adaptiva klassiﬁceringsmodeller.

Målet med detta examensarbete är att jämföra flera olika övervakade klassi- ficerares (eng. supervised classifiers) prestanda när dem tränats med hjälp av ett iterativt användarspecifikt aktivitetsklassificeringsmetod, som beskrivs i denna rapport. En kvalitativ och kvantitativ datainsamlingsprocess tillämpades. Träd- baserade klassificeringsalgoritmerna Decision Tree, Random Forest samt XGBoost testades utifrån specifikt skapade dataset baserade på 21 volontärer, som delades in i tre grupper. Data är baserad på rörelsedata från armbandssensorer.

Beräknat över samtlig data, ökade den genomsnittliga sensitiviteten med 5.2%

(simulerad korsvalidering genom utelämna-en-individ) för algoritmer tränade via beskrivna metoden jämfört med slumpvis icke-iterativ träning.

Nyckelord:aktivitetsigenkänning, övervakade klassiﬁcerares, random forest, xgboost, beslutsträd, iterativt lärometod, användarspeciﬁk

(8)

Acronyms

AdaBoost Adaptive Boosting

ADL Activities of Daily Living

AI Artiﬁcial intelligence

ANN Artiﬁcial neural network

AUC Area under the curve

CART Classiﬁcation and Regression Trees

CNN Convolutional neural network

CPU Central processing unit

CRISP-DM Cross-industry standard process for data mining

DLN Deep learning

DT Decision tree

FN False negatives

FP False positives

GB Gigabyte

GBM Gradient Boosting Machine

GDPR General Data Protection Regulation

GPS Global Positioning System

GR Gaussian regression

HAR Human Activity Recognition

ix

(11)

HMM Hidden Markov Models

HRM Heart rate monitor

IADL Instrumental Activities of Daily Living ICT Information and communication technology

IoMT Internet of Medical Things

k-NN k-nearest neighbor

LDA Linear discriminant analysis

LOF Local outlier factor

LOSO Leave-One-Subject-Out

LSTM Long short-term memory

LR Linear Regression

MGD Multivariate Gaussian Distribution

NaN Not a number

NB Naïve Bayes

NN Neural network

PIR Passive infrared

RF Random forest

RKELM Reduced Kernel Extreme Learning Machine

RNN Recurrent neural network

ROC Receiver operating characteristic

SVM Support vector machine

TN True negatives

TP True positives

XGB/XGBoost Extreme Gradient Boosting

(12)

List of Figures

2.5 A simple decision tree example . . . 18

2.6 Random forest (tree ensemble) example . . . 20

3.3 Windowing example with a 50% overlap from one accelerometer axis 33 3.4 Iterative user-speciﬁc human activity classiﬁcation approach . . . . 36

4.1 Subject-speciﬁc example recording of daily activities . . . 44

4.2 Example recording of daily activities, x-axis only . . . 44

4.3 Example recording of activities walking and standing . . . 45

4.4 Window samples per activity ID per dataset (10.0 seconds approach) 46 4.5 Window samples per activity ID per dataset (1.0 second approach) 46 4.6 Evaluation results dataset_students with 10.0 seconds windows . . . 48

4.7 Evaluation results dataset_students with 1.0 second windows . . . . 49

4.8 Evaluation results dataset_adults with 10.0 seconds windows . . . . 50

4.9 Evaluation results, dataset_adults with 1.0 second windows . . . 50

4.10 Evaluation results, dataset_total with 10.0 seconds windows . . . 51

4.11 Evaluation results, dataset_total with 1.0 second windows . . . 52

4.12 AUC plot of best individual RF run (10.0 seconds window) . . . 54

4.13 AUC plot of best individual RF run (1.0 second window) . . . 54

A.1 One computed example tree from an XGBoost implementation . . . 83

xi

(13)

2.3 Summary of state-of-the-art HAR research . . . 22

2.4 Publicly available HAR datasets (an extract) . . . 24

3.1 Activity script for own data acquisition . . . 29

3.2 Activity recording script . . . 31

3.3 Computed and selected features . . . 34

3.4 Parameter after hyperparameter optimization for the DT implementation . . . 37

3.5 Parameter after hyperparameter optimization for the RF implementation, only modiﬁed parameters are listed . . . 38

3.6 Parameter after hyperparameter optimization for the XGBoost implementation . . . 39

3.7 Empty labelled binary confusion matrix . . . 40

4.1 Window-based best absolute performance increases . . . 53

B.1 Publicly available HAR datasets . . . 85

C.1 Computed features . . . 93

xii

(14)

CHAPTER 1

Introduction

1.1 Background

With a growing number of elderly people in need for additional nursing and health care the integration of additional technologies increases caregiver’s effi- ciency. One specific use case and area of application with continuous increasing interest can be found within the research area of health and nursing care [1]. The deeper understanding of individual user behaviors and patterns are the foun- dation for generating useful user dependent information about personal activity patterns and daily behaviors. This in return can be used to try to predict required actions and improved treatment methods. Therefore, their correct and accurate detections are essential. This accuracy impacts the final usage and can highly differ between various methods. With the deployment of machine learning application the inconceivable amount of manual user profile creations and their maintaining can be intercepted.

With advances in the applications of machine learning algorithms and their steady performance improvement, those algorithms can be used to both classify and predict different activity patterns. Sensor data from for example wearable devices are used as an input source. Besides, further resources like external data information, video material, or direct user feedback can be used to help improving the detection algorithms. Those resulting models and classiﬁers are typically the outcomes of supervised machine learning algorithms which take labeled or categorized preprocessed data as input [2]. An increasing interest in the usage of semi-supervised and unsupervised approaches can be seen as well [3]. Hu- man activity recognition requires user (sensor) data as input for several machine learning based classiﬁer training processes.

An accurate and precise detection of (daily) human activities represents the baseline of several scientific research areas namely physical human activity classification and recognition, sensor data analysis, health care, fitness tracking, and more. They can be categorized by recurring and consistent activities like walking, standing, and/or sitting, by transitional ones like standing up or lying down, and by non-consistent ones like preparing a meal or having a shower. Conclud- ing behavior patterns like having a meal after preparing a meal can be derived from this information in a next step [4]. Besides, the detection of uncontrolled fall movements due to several reasons aligns also with this category.

1

(15)

1.2 Problem

One main problem in the described application area for precise sensor-based activity recognition lies in the ability to identify relevant changes over time without the too cost- and labor-intensive recurring process of collecting and labeling new user data on a per subject basis as a result of changing motion and movement patterns. This is necessary for the creation of learning classification models. Diffi- culties increase for user groups which are not able to actively help or assist in this data collection phase due to several reason, for example evolved aging processes for elderly people. The mentioned problem is followed up and amplified by the lack of reusable classification architectures which are able to adapt changing user behavior over time with the help of different machine learning algorithms.

A significantly high number of studies was done during the last years in the field of human activity recognition (HAR) especially in the area of activities of daily living. Whereas many published researches investigated different sensor data qualities [5, 6] and usages [7, 8, 9], others already clearly showed which machine learning based algorithms potentially work better for which use case [10]. Mannini et al. for example showed that specific algorithms – and especially their fusion – outperformed hitherto classification results [11].

Changing user behaviors and patterns especially during the execution of physical activities require classification algorithms which are able to constantly adapt those changes. This is crucial in the training and learning of new classifiers for new or partly enriched sensor data. Due to the fact of human individuality light differences in the execution of those activities are omnipresent. Therefore, current solutions and approaches have to deal with and find an accepted balance between generalization and personalization what is reflected in the classifiers performances and preciseness. A lack of a more personalized approach which builds on the knowledge of an existing already previously created general baseline classifier is determined. No user engagement and still with the functionality of continuous adaption of movement and motion changes is the primary focus.

To the best of the authors knowledge there has not been an investigation of how and to what extend a novel classiﬁcation approach can help to personalize human activity recognition based machine learning models in an straightforward and reconstructable method with the focus on individual activity class detection.

Those information and insight can be (re)used and analyzed in a next step.

1.3 Purpose

The purpose of this work is to investigate different supervised machine learning classification algorithms applied and evaluated through a novel iterative learning classification approach on a per user basis. This helps to answer the existing knowledge gap of missing approaches to automatically integrate unseen new activity user data into the classifier learning process which then can be used for further processing. This shows the need for a user-dependent classification approach with as little active user interaction as possible. The research question for the available thesis work is therefore defined as follow:

(16)

1.4 Goal 3

‘How and to what extend can a supervised classification approach for sensor-based physical human activity recognition adapt user-specific changing patterns and

behavior?’

1.4 Goal

The underlying goal of this thesis work is to present the comparison of several supervised classifier performances when trained, tested, and evaluated on a novel iterative user-specific human activity classification approach.

This is supported and shown by the identiﬁcation of the most performant comparable tree-based algorithm without the additional effort and need for newly data labeling processes and to provide necessary and evaluated means for its application. Besides, the architectural design of the novel mentioned approach using sensor data for physical human activity recognition is described. The outcome acts as a reusable approach for future and further research on user-speciﬁc HAR.

1.5 Commissioned Work

This work has been commissioned by Noomi AB (formerly Aiﬂoo AB). Noomi AB is a growing startup based and founded in Stockholm, Sweden with the aim of improving the lives of elderly and their caregivers with great respect for personalized integrity. By combining sensor-based hardware with artiﬁcial intelligence (AI) based learning algorithms, Noomi wants to give peace of mind to relatives and simplify the detection of (changing) daily human behavior, falls, and other user behaviors.

No interests of conﬂicts have been found and reported during the development of this work in cooperation with the above-mentioned company.

1.6 Ethics and Sustainability

Ethical issues may be found in personal data collection and usage. Clear and current regulation compliance like correct anonymization policies can prevent those aspects beforehand (demonstrated for example by Cios et al. [12]). A narrow ridge can be found in future user data gathering and the handling of classiﬁed daily activity behavior. That information may result in individual-related data which have to be treated in even stricter ways. Transparency of internal data usage and storage can give corrective. Users need to have the control over their initially collect raw data as well as their transformations. Lastly, provider of applications which use similar kind of systems and frameworks and are dealing with private and personal data receive the obligation of dutiful acting which be- comes clear in for example the ﬁeld of fall detection and prevention [13].

(17)

An uncontrollable self-learning algorithm-based system due to current resources and limitations can be eliminated. The sustainability aspects of the available work can be found in the aimed reproducibility which enables individuals to train, validate, and test own user speciﬁc datasets of different daily activities.

Besides, the developed approach follows a concept with limited maintaining effort and therefore does not require external dependencies, excluding the technical internal implementation. An efﬁcient reuse of both hardware and software components as well as gathered sensor-based user data can be afﬁrmed.

The main targeted audiences for this thesis are alike research individuals and groups and interested readers within the field of activity recognition approaches and a great interest of novel approaches for transforming generalized classification models towards approaches being more able to adapt individual user patterns. They benefit from it in terms of the investigated evaluation performance which can be used for further hypothesis validations. One group of advanta- geous users may also be research departments in companies and institutions within the field of applied HAR for different user groups.

1.7 Methodology

In order to align with the described problem definition, a combined workflow methodology was chosen to merge a fusion of qualitative and quantitative data gathering with an inductive theory creation, which will be described further in Chapter 3, and a theoretical study outline. The reason to choose this approach was based on the required data characteristic whereas the quality and quantity of a dataset are essential for an accurate classification model comparison.

This highly aligns with the cross-industry standard process for data mining (CRISP-DM) model approach [14]. The data gathering was done with the help of healthy volunteering activity performing subjects per deﬁned activity class.

Furthermore, an inductive approach was chosen as the focus of this work is not the creation of new theories based on newly generated data rather than the goal- oriented usage of it with the described outcome. This was done by several ex- periments using the collected datasets with aiming performance comparisons. A more detailed description of the underlying followed methodology can be found in Chapter 3.

1.8 Delimitations

The work carried out focused on daily activities performed by healthy human volunteers using and analyzing the sensor data input of two body-worn sensors.

The investigated machine learning approaches belong to the group of supervised tree-based techniques. The focal point was the performance comparison based on ﬁnal averaged true activity class detections applied on two different sliding time window concepts. Explicitly aspects and features which were not developed are mentioned, discussed, and reasoned in Section 4.4.

(18)

1.9 Outline 5

1.9 Outline

Chapter 2 introduces a necessary deeper background in the ﬁeld of human activity recognition and related research, followed by the project’s methodology in Chapter 3. The resulting evaluation including the comparison between several machine learning approaches are presented in Chapter 4, which are followed by evaluations and result transferability discussions in the following Chapter 5.

The ﬁnal Chapter 6 reﬂects the work done and opens space for further work and points of reference.

(19)

(20)

CHAPTER 2

Elderly Care, Activity Recognition, and Technological Support

In this chapter a detailed description about relevant background sections for this degree project is shown together with related work and resources. Section 2.1 describes the scope of the application which includes healthcare, elderly care, and potential usages of new technologies in those ﬁelds followed by Section 2.2 which describes the basic movement and motion sequences of a human body. Those described contents are essential for the experimental methodology in Chapter 3 which puts the underlying use case in context. Section 2.3 highlights the components of physical activity recognition as well as different iterative learning approaches. They give insights about the experimental execution in Chapter 3. Sec- tion 2.4 describes the chosen algorithm approaches and their functionality and Section 2.5 closes this chapter with relevant and related work. Those are highly relevant for a better understanding of this work’s results (see Chapter 4) and evaluations (Chapter 5).

2.1 Scope of Application

This section shows an introduction to the ﬁeld of health care in general (see Sec- tion 2.1.1), its application for elderly in particular (see Section 2.1.2), and possible points of reference with technological usages, services, and applications in elderly care (see Section 2.1.3).

2.1.1. Health Care

Health care is the ‘[...] organized provision of medical care to individuals or a community [...]’ [15]. The World Health Organization (WHO) extends this term to primary health care which ‘[...] is about caring for people, rather than simply treating speciﬁc diseases or conditions [...]’ [16] and which shows an increasing interest and attention.

As health care is provided by different entities and in different ways, the main aim for health care systems is the improvement of the health of covered caretak-

7

(21)

ers. Differences in resulting cost structures and health outcomes can be observed between different countries, religion, cultures, and underlying providers [17].

The expenditures for health care have been growing globally over the at least 15 to 20 years. The growth of per-capita spending in health care can be split into age and sex factors, medical prices, and residual growth components which are attributable to use and intensity. Those attributes entail the different characteristics within the ﬁeld of health (care) insurances [18]. The itemization of those exceeds the scope of this work.

2.1.2. Insights Elderly Care

Growing populations around the world are ageing with an increasing speed.

Ageing in this context represents a natural process and applies to all sections of a society. The entry age of the age class of elderly can be deﬁned as between 60-65 years and can be based on individual well-being, origin, development, and human characteristics as well. Therefore, it is not possible to determine it exactly [19]. By the year 2025, the group of elderly person will grow up to 1.2 billion and respectively up to around 2 billion by the year 2050 [20] globally. This states an increase of around 208%, considering around 960 million in the year 2017. The same trend can be seen in the age grouping determined over 80 years which will approximately triple from around 140 million in 2017 [21].

With a growing age, the possibilities to suffer from chronic diseases, physical disabilities, and mental illnesses increase which highly correlates with a higher demand for the need of elderly care [22]. The potential prevention and active control of increasing health issues of elderly consists of a multifaced concept combining the collaboration of health and healthcare, legal sectors and limitations, and wide-ranging social welfare [21]. The functional ability for older adults depends on individual health conditions and lifestyles. Especially the transition from work to retirement can cause a signiﬁcant change in the amount and execution of daily activities which then effects the personal health status [23].

The process of taking care of elderlies, also referred to as eldercare, is carried out in several forms and use cases. They range from independent living commu- nities, assisted living facilities, over nursing homes (also referred to as residential living), aging in home/home care approaches, to continuing care retirement com- munities, and end-of-life care [24, 25, 26, 27]. Two attributes within this ﬁeld of caretaking are the tools and incoming equipment to ensure a controlled, efﬁcient, and functioning nursing and care. Different care providing approaches require different tool sets which at the same highly depend on individual use cases and health statuses and progress.

Falls, which account for a significant high number of injuries for elderly, are defined as ‘[...] inadvertently coming to rest on the ground, floor or other lower level, excluding intentional change in position to rest in furniture, wall or other objects [...]’ [28]. Various researches estimated the percentage of people aged over 65 years having at least one unintended fall per year up to 30%, followed by around 50% for people aged 80 years or older [28]. The fall risk is thereby higher in nursing homes than in community living facilities. Falls cause around 40% of all injured deaths per year. The main reason for falls can be found in bi-

(22)

2.1 Scope of Application 9 ological risk factors (for example age, gender, race), behavioral risk factors (for example intake of medicine or alcohol), environmental risk factors (for example insufﬁcient lighting or slippery groundings), and socioeconomic risk factors (for example income, education, housing). Most falls and their resulting injuries are the result of a combination of age, individual interaction with the physical and social environment, and possible already existing disease-related factors of individuals [29]. This kind of health-related risks can cause many different kinds of impacts and short and long-term results and therefore gained a growing attention over the last years and decades.

Regarding the correct and in-time fall detection and prevention as a previous step many health care and social service providers are not prepared for an accurate managing of those issues. One non-technical approach for an effective fall prevention is suggested by the WHO in their Active Aging Framework which describes proactive and ﬂexible prevention methods. It combines education, capacity building, and training in a community environment with a healthy public policy via surveillance, resources, and research [29].

2.1.3. Technology in Elderly Health Care

Health and wellbeing for seniors can be improved and at the same time made easier for caregivers by preconceived integration and usage of new services and applications mainly based on technological improvements and renewals [30, 31, 32, 33]. The process of creating synergy between those two areas can look like following: i) knowledge and user dependent data extraction with the help of sensors¹and supervision tools, ii) user behavior and activity recognition, iii) processing of the gained knowledge for machine and computer interpretation and post- processing, iv) the usage of intelligent algorithms for complex pattern recognition and reasonings, and v) the right way of communicating the right information between caretakers and -givers [34].

The need for an increasing efﬁcient usage of information and communication technology (ICT) in the sector of elderly care is encouraged by analyzing the reasons for elderly not accessing health care services and facilities. Table 2.1 shows a summarized reasoning analysis and it can be highlighted that the fact that potential patients were not sick enough is an important and consistent reason throughout all income segments.

State-of-the-art trends and developments in the active usage and integration of technology in the health care ﬁeld and market of elderly can be found in care

1Sensors in this context can be electronic, optical, thermal, chemical, or genetic sensors, including correlating vital measurements like pulse, cardiogram, body temperature, oxygen level, blood pressure, and more [34].

Table 2.1.Reasons for not accessing health care services (adapted from [28])

Reason Country income category (% of respondents)

High income Upper-middle-income Lower-middle-income Low-income

Could not afford the visit 12,1 19,3 20,7 29,1

Health-care provider’s skills inadequate 19,0 8,3 7,8 13,1

Did not know where to go 12,2 9,7 9,8 7,8

Was not sick enough 21,5 31,8 27,3 25,8

(23)

robots, the Internet of Medical Things (IoMT) which connects all integrated digi- tal devices for an advanced data analysis, the utilization of virtual reality systems, predictive and preventive analysis based on individual (customer) datasets, and telemedicine in various forms and delivery options [35].

Current more concrete examples of new solutions and services with a very high technology correlation are the ﬂexible in-house care professional booking system Honor [36], the startup Virta health which aims to reverse diabetes and offer patients personalized treatment plans and remote doctor appointments [37], and the pairing platform Hometeam which aims to match carers and aides with seniors regarding their overlapping interests [38].

Especially mobile health applications for elderly are emerging during the last years which combine mobile communication devices with mainly wearable sens- ing technologies. The entry of those scalable health technologies enable user- dependent data collections [23]. The adoption rate of new ICT based solutions for both parties, caregivers and -takers depends on external factors like the ease of use, the device performance, the non-involvement of the user in the selection of the technology, and the evolving user priorities [39].

2.2 Human Body Movements and Motions

The complexity of the human body and its correlated movements and motions can be best described with the help of defined anatomical terms of motions. The reason for a wide range of human body movements lie in the capability of performing synovial joints. Therefore, each movement or synovial joint is the result of either a contraction or relaxation of connected muscles. Several combined or sequential joints of this kind produce a specific body movement and includes several types and amounts of synovial joints per execution [40]. In the following the main anatomical motions are described which contribute and occur among the most within the field of human motion sequences:

• Flexion: Describes the process of the decrease of an angle between a body part and its proximal segment (for example bending a knee or a forearm). It especially deﬁnes the action of bending a limb or joint, but is not limited to those [41].

• Extension: This process describes a movement that increases the angle between different parts of the body. Both, ﬂexion and extension, are counter- parts and occur together in most of the use cases [42].

• Abduction: Describes the movement which moves a part or segment (for example an arm or leg) of the body away from its midline. The midline in this context represents the physical center of a human body. One example can be found in the spreading of ﬁngers as this implies moving apart from a physical, in this case a hand [43].

• Adduction: The countermovement to abduction and therefore the movement that pulls a body part towards the centerline of a body or another body part [43].

(24)

2.2 Human Body Movements and Motions 11

Figure 2.1.Two examples of ﬂexion and extension [44]

• Rotation: Describes the movement of rotating a part of the body towards (internal) or away the center axis of a body (external). It is very often the result of the summation of several small rotational movements [43].

• Circumduction: Describes the sequential combination of above mentioned ﬂexion, adduction, extension, and abductions movements of the limb, hand, or ﬁngers. Thereby, one end of a body region which moves stays relatively static in this movement while the other end performs a circle [40].

Besides the above mentioned, academic literature mentions elevation (for example an upward movement of the shoulder and scapula), depression (down- wards counterpart of elevation), pronation (for example a movement that changes the forearm from a supinated to a pronated position), supination (counterpart of pronation), inversion (for example turning a foot to angle the bottom), and ever- sion (for example turning a foot’s bottom facing the ground) [40], and more body movements within this section. Figure 2.1 illustrates an example of ﬂexion and extension in the use case of a simple human upper body movement as they are comparably often executed individually.

As mentioned earlier and to deﬁne a placeholder for the upcoming sections an averaged adult human body conﬁguration is used which symbolized an ex- perimentee with normalized and mean attributes. Table 2.2 shows the main attributes. It has to be noted that the used data was gathered from American males

Table 2.2.The standard human (adapted from [45])

Attribute Value

Age 30 years

Height 1.72m (5 ft 8in)

Mass 70 kg

Weight 690 N (154lb)

Body surface 1.85m²

Body core temperature 37.0°C Body skin temperature 34.0°C

Basal metabolic rate 70 kcal/h (1,680 kcal/day)

Body fat 15%

Heart rate 65 beats/min

(25)

in the mid-1970s and that average values may change over time for different re- gions, countries, as well as genders [45].

A speciﬁc human body movement like the periodic activity of moving/walking forward for a speciﬁc period by using both legs and several parts of the lower human body is the combination of several above-mentioned anatomical motions.

Besides these examples different combinations of different anatomical body part movements result in universal recognizable movements and activities like walking, running, standing, or others.

2.3 Physical Activity Recognition

The following sections highlight the key principles of physical activity recognition. The importance of human physical activities is described in Section 2.3.1, possible sensor conditions for an efﬁcient activity tracking in Section 2.3.2, dif- fered methods for the recognition of those activities in Section 2.3.3, and an introduction to related iterative learning approaches in Section 2.3.4, respectively.

2.3.1. Physical Human Activities

The learning and accurate detection of human activities highlights a promising step towards improved elderly care which is driven by growing health care in- dustries, increasing technological progress, and the demographical changes in a growing number of countries [30, 34]. HAR is used to serve a large field of applications ranging from health and elderly care to personal fitness, indoor naviga- tion, emergency services, gaming, security surveillance, and creative applications scenarios. External or wearable sensors serve as input resources and generate user-dependent sequences of time series based data samples. Different activities and movements of different users can then be classified by generated features which try to capture statistical differences from those data samples.

Kinematics is the study of human motion. Newton’s first law relates to that which defines that an object in rest stays in rest whereas an object in motion stays in motion. (Human) Motion in this context is split into translatory motion (all parts of a moving body move towards the same line or direction) and rotational motion (the motion of rotation of a moving body or parts of it). Both combined define the category of general movements and motion [46]. To achieve this, skele- tal body muscles which require energy expenditure produce any kind of bodily movement. Those movements in a defined order and over a specific period of time are defined as physical movements [47].

Recent studies published that physical activity is no longer a direct require- ment for individual daily living. The formerly stronger relationship between physical work and food intake has been suspended [48].

One grouping within the field of physical movements is represented by activities of daily living (ADLs). They therefore describe the ability of daily self- care and can be used as a measurement for the survivability status of a person, especially in the field of healthcare and eldercare [49]. They are not fixed and

(26)

2.3 Physical Activity Recognition 13 standardized. The following Graphic 2.2 shows a summarized overview of generalized daily activities with a cross categorization of instrumental activities of daily living (IADLs). IADLs describe a group of more complex activities which are necessary for a daily self-care and fundamental viability. They are overlapping with ADLs and in many cases build up on top of them [50]. One example of IADLs in comparison to ADLs would be the activity of self-controlled meal preparation or managing medications.

Besides the more speciﬁc groups of activities above, Ainsworth et al. categorized human activities in more general clusters, namely into bicycling motions, lawn and garden, sports, conditioning exercises, transportation, dance and mu- sic, walking activities, ﬁshing and hunting activities, occupation, water activities, home activities, running, activities, home repair, self-care, religious activities, in- activity, sexual activity, volunteer activities, and miscellaneous [51].

2.3.2. Sensor Technologies

Sensors measure and/or detect physical characteristics and reply to them via recording, indicating, or other responses [52]. Widely used sensors in the ﬁeld of activity recognition can be split into two groups. First, physical sensors which are attached to parts of the body of humans or objects. They include accelerome- ters, gyroscopes, temperature sensors, barometers, as well as light and ultrasonic sensors. Second, video and picture based capturing sensors which remotely ob- serve and track the movements of human and objects. This group includes passive infrared (PIR) sensors, Bluetooth, microwave and ultrasonic sensors, video cameras and microphones, as well as tomographic motion and global positioning system (GPS) sensors [53].

The application area for physical sensors range from healthcare and assisted living, over prevention of diseases and early detection of health issue indicators

Figure 2.2.An extended diagram of activities of daily living (ADLs)

(27)

Figure 2.3.Relative x, y, and z orientations of an accelerometer sensor [56]

until the supportive usage in promoting human exercise. Within this range the video-based category covers behavioral biometrics recognition, security and surveillance aspects, interactive applications, and content-based video analysis [53].

From the mentioned application areas above accelerometer and gyroscope sensors are among the mostly used ones in the ﬁeld of activity recognition [53, 54].

An accelerometer is a sensor which measures the rate of change of the velocity of the movement of an object, the acceleration. Acceleration is defined as force divided by mass. The measurement unit is defined as meters per square second (m/s²) or equivalent to G-forces (measured in g). 1 g on the planet Earth is defined as 9,81 m/s² (wherefor an accelerometer during a free fall indicates 0 g). The smaller the measurable range of an accelerometer, the more sensitive the output values. To measure either static or dynamic acceleration this electrome- chanical sensor uses either one, two, or three axes (3-axis or triaxial). These axes indicate the movement direction and are generally labeled as x, y, z which indicate the three axes in a three-dimensional space or system. Internal movable capacity plates create changes in the capacitance between them from which the acceleration can be calculated [55]. Figure 2.3 shows a labeled axes example on a three-dimensional object.

A gyroscope on the other hand is a sensor which measures angular velocity and therefore rotational motion. Angular velocity is deﬁned as the speed of rotation. The measurement unit is degrees per second (°/s) or also revolutions per second. Either one, two, or three axes (3-axis or triaxial) are used. These axes indicate the rotation direction per axes and are generally labeled as x, y, z which indicate the three axes in a three-dimensional space or system as well. These mi- croelectromechanical systems use an internal resonating mass to detect angular velocity changes by its shifting movement, transfer this movement into electrical signals, and therefor identify the change [57]. Figure 2.4 shows a labeled axes example on a three-dimensional object.

Hybrid or combined sensor approaches are present if at least two approaches (sensor based as described above, vision based, electrogoniometer based or mag- netic and medical imaging-based systems) are combined. Depending on the application use case they result in better performances compared to single-sensor

(28)

2.3 Physical Activity Recognition 15

Figure 2.4.Relative X, Y, and Z orientations of a gyroscope sensor [59]

solutions [58]. The description of further sensors as well as different sensor fami- lies exceeds the scope of this work.

2.3.3. Activity Recognition Methods

Several methodologies, concepts, and frameworks are used within the ﬁeld of human activity recognition, depending on the use case and balance between implementation complexity and desired outcome. All aim to recognize common human activities in real life as accurate as possible. This is done via the generation of feature sets over the captured data.

In general, the distinction between online and offline activity recognition approaches should be made. Whereas online learning implies the detection of human activities in real-time and with present and continuous data input, offline approaches deal with in-laboratory settings and focus on testing and development purposes first. More challenges are faced within the group of online approaches. This helps to clarify the aimed outcome of a problem statement as well as the followed workflow and delimitations.

One possible methodology for the detection of different activities is based on predefined logic rules and rule sets. Rules in this context can define the name of an activity, the minimum and maximum time of an activity, latency periods, peak thresholds, or minimum in-between time ranges (=measurable facts). The defined rules are stored in a centralized database whence different applications can request that information. The number of manually created rules for simple ADLs can easily exceed an amount of 90 and more per activity. Storf et al. besides showed one successful example of a working rule-based activity recognition approach [60]. A next step in the development of rule-based detection and learner systems is the automatic generation of rules based on gained knowledge from similar or previous activities. Fuzzy logic is one rule-based approach applicable for HAR as it handles vagueness, insufficient knowledge, and imprecision. It is used in systems where data points and observations cannot be represented in a binary form [58].

Besides rule-based learning techniques, evolutionary learning within the ﬁeld of physical motion recognition plays an increasing role. Genetic algorithms, as

(29)

one extract, can be used especially for optimization-based problems as their functionality is inspired by biological evolution. This optimization can be applied to the preprocessing step of selecting features for a motion recognition algorithm and to selecting the optimal classiﬁers [58].

Probabilistic modelling is used in ﬁelds with higher uncertainty and data noise with the help of mathematical probability calculations. Possible techniques are Bayesian statistics, Hidden Markov Models (HMM), or Gaussian regression (GR) [58]. Combining above mentioned techniques as for example a rule-based approach with a neural network (NN) architecture, hybrid learning can be applied for better feature recognitions. At the same time, supervised and unsupervised techniques can be combined to follow a semi-supervised approach [58].

Another area within the ﬁeld of unsupervised learning techniques covers the ﬁeld of NNs and especially convolutional neural networks (CNNs), recurrent neural networks (RNNs), deep neural network architectures or combinations and/

or modiﬁcations of the above mentioned.

2.3.4. Iterative Learning Approaches

The history of iterative ways of solving (mathematical) problems dates back until a year around 1823 where the mathematician Carl Friedrich Gauss proposed the usage of iterative methods for solving speciﬁc systems of linear equations to his students [61].

In general, an iterative approach describes the methodology of reusing gained information from an initial research interval and session in a repetitive sequence with the aim of converging as close as possible towards a predefined measurable goal [62]. Therefore, the approximation of x_n+1is gained from the approximation of x_n. This definition can be translated to the field of mathematical interrogations.

A simple representation of this can be seen in equation 2.1.

x_n+1= f(x_n) (2.1)

Note that the term iterative can be replaced and exchanged bidirectional with recursive at this point.

x1 =x0− ^f (x0)

f^′(x0) ^(2.2)

Carrying on, equation 2.2 shows the Newton’s method, an recursive well- known example for ﬁnding better approximations to roots of a real-valued function [63]. The process of this iteration is repeated until a result sufﬁciently accurate can be achieved.

A high percentage of application and acceptance can be found in the field of advanced mathematical problem solving as well as machine learning. One of their main strengths is the scalability regarding large amounts of input data which shows their fortes compared to more fixed linear solution approaches and algorithms. It is therefore not essential to set initial parameters for the initial iteration as an efficient algorithm architecture finds a goal-oriented value optimiza-

(30)

2.4 Algorithm Concepts 17 tion in the first iterations without external influences. Considering the training and testing phases of machine learning based models, an iterative solution starts with an initial guess for input weights and parameters and iteratively adjusts those first guesses [64].

2.4 Algorithm Concepts

This section describes the theoretical background of the operating principles of three tree-based supervised machine learning algorithms. The concepts of decision trees, random forests, and extreme gradient boostings are explained in Sec- tion 2.4.1, Section 2.4.2, and Section 2.4.3 respectively. They all have been used, tested, and evaluated in the experimental setup described in Chapter 3.

2.4.1. Decision Tree

Decision tree (DT) based machine learning concepts belong to the field of supervised learning approaches. They can be used for classification and regression use cases and work with categorical and continuous output and input values. Differ- ent underlying algorithms which define the internal functionality and logic exist, as for example the widespread Classification and Regression Tree (CART) algorithms [65]. For a given input value x, the root of the tree compares this value with a determining root variable x1. If x is smaller or equal to x1, the tree continues the creating process of a child node at this level on the left hand side, if x is bigger than x1, on the right hand side. This process is defined as a tree split and is repeated until all predefined determining variables (also features) are queried, so that there are no more non-terminal nodes which could cause another split. A terminal node indicates the last step of one process cycle and assigns x a related result or label y1. Y1 in this case is defined as the leave of this path of the tree.

Figure 2.5 shows a simple decision tree example with imaginary values and three features and four target classes. Parameters and variables used within the context of decision trees are mentioned and described as follow [66, 67]:

• Maximum depth: indicates how deep a tree can grow. The depth correlates with the number of splits. More splits cause a deeper ﬁnal tree structure which then contains more information.

• Minimum samples per split: determines the minimum number of input samples which are required for a node to be split. A higher number cre- ates a more meaningless and limited tree.

• Minimum samples per leaf: determines the minimum number of input samples which are required for a ﬁnal leaf node.

• Maximum number of features: indicates the number of input features (of all samples) which is considered for creating a new split.

• Maximum number of leaf nodes: deﬁnes the number of leaf nodes which contain the ﬁnal leaves.

(31)

Figure 2.5.A simple decision tree example

• Minimum impurity for splitting: determines a ﬁxed threshold for the early stopping of a tree. Tree growing continues if a node’s impurity is above this threshold.

• Criterion: determines the underlying function for measuring the split criteria of a node. Commonly used approaches are based on the Gini Index or the Information Gain (see below).

• Splitting strategy: possible splitting strategies can be based on the best value or based on a random approach.

Different settings can influence the model performance in different peculiari- ties. Each learning sample of a decision tress consists of a set of N cases. Each of those cases is defined by at least one measurement vector and a resulting class or category j it belongs to. One learning sample therefore is defined as in equation 2.3.

L= {(x_k, j_k)}_k^N₌₁ (2.3) L(t) therefore defines a set of associated cases to node t. The construction process of a new tree takes the root of the tree, for example to which all input learning samples of L(t₀) = Lare assigned. If all input cases at this step belong to the same class j, the splitting process is done and t0 is defined as a/the leaf of this path. If not, the cases in L(t)are split into two sets. The benefit of a possible split and its reason for being created have to be evaluated. Therefore, I(t) as a quality variable per split is defined to measures the impurity of all cases in L(t). I(t) = 0 (= pure) is true if all cases in L(t) belong to the same class. Splitting increases the impurity of a node. One very widely used approach is the Gini measure or index as found in equation 2.4

I(t) = 1−

∑

^P⁽^j^|^t⁾² (2.4) P(j|t)therefore describes the probability that a case in L(t)belongs to the class j. The maximum impurity is achieved by equally represented classes in L(t). The

(32)

2.4 Algorithm Concepts 19 maximum number of splits which are considered during the construction process of a new tree equals the number of variables times the number of different classes in L(t). All splits are taken into account during one building process and the split which maximizes the change of the impurity is defined as the best for this node [68]. Therefore, the goal of a split is to find the most significant variable from the input set and its corresponding value which generates the best homogenous sets of samples. Information gain on the other side is defined as the result from one minus entropy whereby entropy is defined in equation 2.5.

entropy = −plog₂p −qlog₂q (2.5) P and q are placeholders for the probability of success (input value has been classiﬁed correctly) and failure (input variable has been classiﬁed incorrectly) in one candidate splitting node.

Two main techniques can be used to perform ensemble DTs and to move one step forward of combining several DTs. First, bagging generates random sub- samples of the training input samples with replacement and trains one decision tree per subsample. The average of all predictions is used as the ﬁnal output which can result in a more robust outcome and a reduced variance of this DT.

Second, boosting builds models/DTs sequentially instead of independently. A build predictor learns from the mistakes and errors made from the previous one and can adjust the new construction. Therefore, this technique results in faster ﬁnal predictor ﬁndings and is used to produce a result with reduced bias [69].

One overall approach of improving the performance of DTs and for reduc- ing the risk of overﬁtting trained models is pruning. Thereby, parts of the built tree that provide too little impact and power of classifying input samples are re- moved. This cutting back process can be done by several strategies, for example cutting the smallest tree leaves or nodes or cutting the ones with the smallest error measure [70].

2.4.2. Random Forest

The random forest (RF) algorithm is an example of an ensemble machine learning method and belongs to the group of non-parametric algorithms. Ensemble methods act as aggregators which use a voting procedure, like majority voting, to determine the ﬁnal outcome among a set number of different estimators [71].

Those estimators subsist of several decision (and also classiﬁcation) trees whereas those build the basic constructing block for a RF implementation.

A random forest algorithm builds several DTs and takes the advantage of their averaged meaningfulness and information support for a predicted outcome. It can be used for both regression and classiﬁcation problems. The ﬁnal voting step in a RF structure can be based on different strategies, for example average cal- culation. Besides, each DT taken into account in a RF structure consists of both a random subsample of the whole input training dataset (application of the bagging approach) and a random subsample of the indicative features per newly generated input sample set, both with replacement. This differentiates a RF from a bigger bagged DT.

(33)

Figure 2.6 shows a simple majority voting approach of a RF with ﬁxed parameter settings. The following parameter settings where chosen: i) maximum depth

= 2, ii) maximum number of features per split = 1, iii) minimum samples per leaf

= 20, and iv) minimum samples per split = 50.

Possible hyperparameters hereby overlap with the list described ones from DTs. Besides, RF implementations offer more possible tuning parameters like the total number of estimators (deﬁnes the number of to be build DTs using a provided set of input data samples), the possibility of a warm start (reusing a solution of a previous run and adding and ﬁtting more estimators to ensemble) or the usage of an out-of-bag score (including out-of-bag samples for an estimation of the generalization accuracy).

The aim of RF implementations is to make the created DTs de-correlated and use pruning strategies by setting stopping criteria for new node splits. This in general helps a RF structure to prevent model overﬁtting which is more likely to occur in plain DT systems. By using many subsample-based binary classi- ﬁcation DTs as a baseline this algorithm provides useful information about the importance of each individual input feature in context of a model performance in general. This is due to the fact that at each node of newly build DT a given number of input features are chosen randomly and that the best split for possible children nodes is based on only this subset [72]. Whereas a RF implementation takes comparable more computational resources and execution time than a DT, it can execute the training phase and therefore the building step of the underlying DTs in parallel.

2.4.3. Extreme Gradient Boosting

Extreme Gradient Boosting (XGBoost) is based on the origin model of a gradient boosting machine (GBM), also known as gradient tree boosting or gradient boosted regression tree and belongs to the group of tree ensembles. Boosting systems provide state-of-the-art outcomes to many classiﬁcation problems. XG- Boost therefore implements a gradient decent based boosting algorithm for decision trees. In many use cases it outperforms boosted tree architectures with the main advantage of using a minimal amount of resources and faster simultaneous execution times. Besides, it introduces more randomization and regularization options, namely L1, L2, and penalization on the amount of leaf nodes [73, 74].

Figure 2.6.Random forest (tree ensemble) example

(34)

2.5 Related Work 21 This sparsity-aware algorithm focuses on sparse data as well as weighted quantile sketches for better approximate tree learning. The most important factor of XBoosts growing success rate and research attention is its ability to scale in memory-limited or distributed settings. It exploits out-of-core computation to increase the processing time even more [74]. An example illustration from one computed decision tree from an XGBoost implementation, generated during the experimental test runs in Section 4.2, can be found in Appendix A.

2.5 Related Work

The following related work gives an overview of current and state-of-the-art researches in the field of human activity recognition and classification model creations for sensor data as combination of all five previously described relevant background fields.

Regarding the in Section 2.3.3 mentioned approaches for identifying and de- tecting daily human activities including falls as a subcategory, a solid amount of research in this area has been done with noteworthy results for different application areas. The following Table 2.3 summaries exceptional resulting and state-of- the-art ones using the abbreviations acc. (accelerometer), gyr. (gyroscope), mag.

(magnetometer), t. (time domain), and f. (frequency domain).

An increasing interest regarding the in Section 2.3.4 mentioned methodology for iterative algorithms and approaches resulted in many publications and result demonstrations with different focuses and optimization aims. Zhao et al for example presented an user dependent approach, combining K-means clustering algorithms and decision trees with a focus on mobile phone collected accelerometer data and the amount and size of retraining iterations [83]. Viet et al. showed a combination of Support Vector Machine a (SVM) classiﬁer with K-medoids clustering for personalized activity models to recognize even new activities which were not present and available during the trained phase of the base classiﬁer. The SVM training step showed comparable high training costs [84]. Fallahzadeh et al.

showed a transfer learning approach and a novel method for identifying highly similar automatically labeled samples. They followed the purpose of creating comparable small new user-dependent training datasets at the expense of performance losses [85]. Deng et al. focused on the application of RKELM (Reduced Kernel Extreme Learning Machine) for creating user speciﬁc activity recognition models [86].

With the focus on an efficient and result increasing approach, Wen et al. proposed a approach for integrating new sensor devices during the training phase in combination with an AdaBoosting (AdaBoost) algorithm [87]. Zhao et al. recently (as of early 2018) published their work for finding the most profitable previously unseen samples from a new target user/subject with the help of K-means clustering and a combination of local outlier factor (LOF) and multivariate Gaussian distribution (MGD) [88].

Several other authors within the ﬁeld of personalizing classiﬁcation models for activity recognition of different motions groups focused more on subsegments and preprocessing stages for example Li et al. by comparing different methods

(35)

Table2.3.Summaryofstate-of-the-artHARresearch

Sourcerefer-ence TitleClassiﬁcationmethodology Sensors(place-ments) Parametersset-tings EvaluationresultFeaturestrategy(tvsfdomain) Year#ofSub-jects #ofActiv-ities UseddatasetSunetal.[75]Activityrecognitiononanaccelerometerembed-dedmobilephonewithvaryingpositionsandorientations SVMAcc.10Hz94.8%F1-scoreT.&F.201077Own Tomasetal.[76] Activityrecognitionfromaccelerometerdataonamobilephone k-NNAcc.30Hz90%accuracyT.2009-6Own

Manninietat.[77] ActivityRecognitioninYouthUsingSingleAc-celerometerPlacedatWristorAnkle SVMAcc.(wristandankle) 90Hz88.5%(wrist)and91.6%(ankle)accu-racy T.&F.20185326/23Own Trostetal.[78] Artiﬁcialneuralnetworkstopredictactivitytypeandenergyexpenditureinyouth ANNAcc.(hip)30Hz88.4%accuracyT.201210012Own

Hagenbuchneretal.[79] Predictionofactivitytypeinpreschoolchildrenus-ingmachinelearningtech-niques ANNanddeeplearning(DLN) Acc.(hip)100Hz69.7%(ANN)82.6%(DLN)accuracy T.20151112Own DelRosarioetal.[80] Acomparisonofactivityclassiﬁcationinyoungerandoldercohortsusingasmartphone DecisiontreeAcc.,gyr.,bar.(trouserspocket) 10010(bar.)Hz80.95%accuracy(average) T.2014579Own

Trostetal.[81] Machinelearningforactiv-ityrecognition:hipversuswristdata Logisticregres-sion Acc.(wristorhip) 3Hz91.0%(hip)88.4%(wrist)accuracy T.20145212Own Chowdhuryetal.[82] EnsembleMethodsforClassiﬁcationofPhysi-calActivitiesfromWristAccelerometry EnsembleMethods(BDT,k-NN,SVM,ANN) Acc.(wrist)32/33Hz76.88%/85.99%F1score T.&F.20178/175/12Own(andPAMPA2)

A Study of an Iterative User-Specific Human Activity Classification Approach

A Study of an Iterative User- Specific Human Activity

Classification Approach

NIKLAS FÜRDERER

School of Electrical Engineering and Computer Science Kungliga Tekniska Högskolan

A Study of an Iterative User-Specific Human Activity Classification Approach

M

T

EIT ICT Innovation and Data Science

Author: Niklas Fürderer

Academic Supervisor: Henrik Boström

Industrial Supervisors: Arnrun Thorsteinsdottir & Tor Carlsson Examiner: Mira Kajko-Mattsson

Course 2017-2018

Acknowledgements

Abstract

Sammanfattning

Contents

Acronyms

List of Figures

CHAPTER 1

Introduction

1.1 Background

1.2 Problem

1.3 Purpose

1.4 Goal

1.5 Commissioned Work

1.6 Ethics and Sustainability

1.7 Methodology

1.8 Delimitations

1.9 Outline

CHAPTER 2

Elderly Care, Activity Recognition, and Technological Support

2.1 Scope of Application

2.1.1. Health Care

2.1.2. Insights Elderly Care

2.1.3. Technology in Elderly Health Care

2.2 Human Body Movements and Motions

2.3 Physical Activity Recognition

2.3.1. Physical Human Activities

2.3.2. Sensor Technologies

2.3.3. Activity Recognition Methods

2.3.4. Iterative Learning Approaches

2.4 Algorithm Concepts

2.4.1. Decision Tree

∑

2.4.2. Random Forest

2.4.3. Extreme Gradient Boosting

2.5 Related Work