ORIGINAL ARTICLE

Journal Section

Recent Trends in Machine Learning for Human

Activity Recognition - A Survey

Sreenivasan Ramasamy Ramamurthy1∗ | Nirmalya Roy1∗

1 Department of Information Systems,

University of Maryland Baltimore County, There has been an upsurge recently in investigating machine
Baltimore, Maryland, 21250, USA learning techniques for Activity Recognition (AR) problems
Correspondence as that have been very effective in extracting and learn-
Department of Information Systems, ing knowledge from the activity datasets. The techniques
University of Maryland Baltimore County,
Baltimore, Maryland, 21250, USA ranges from heuristically derived hand-crafted feature-based
Email: {rsreeni1, nroy} @umbc.edu traditional machine learning algorithms to the recently de-
Funding information veloped hierarchically self-evolving feature-based deep learn-
This work is partially supported by the ing algorithms. AR continues to remain a challenging prob-
Office of Naval Research Grant
N00014-15-1-2229 and the Alzheimer’s lem in uncontrolled smart environments despite the amount
Association Research Grant of work contributed by the researcher in this field. The com-
AARG-17-533039.
plex, volatile, and chaotic nature of the activity data presents
numerous challenges which influence the performance of
the AR systems in the wild. In this article, we present a com-
prehensive overview of recent machine learning and data
mining techniques generally employed for AR and the under-
pinning problems and challenges associated with existing
systems. We also articulate the recent advances and state-
of-the-art techniques in this domain in an attempt to iden-
tify the possible directions for future activity recognition
research.

KEYWORDS
Activity Recognition, Data Mining, Machine Learning, Transfer
Learning, Deep Learning, Active Learning, Wearable Sensors

Abbreviations: AR, Activity Recognition; HAR, Human Activity Recognition.

∗ Equally contributing authors.

1
2 RAMASAMY RAMAMURTHY ET AL .

1 | INTRODUCTION

Extracting knowledge from the raw data, in general, has provided useful information in various fields. Human activity is
unique, as the information inferred from raw activity data has been proved to be critical in functional and behavioral
health monitoring (activities of daily living, sleeping, eating etc.), game console designing, personal fitness tracking,
sports analytics to name a few. Data mining and machine learning approaches have proven to be effective than
the classical mathematics and statistical techniques in extracting knowledge and discovering, learning and inferring
activity from data. Human Activity Recognition (HAR) refers to the automatic detection of various physical activities
performed by people in their daily lives. A HAR system helps recognize the activities performed by a person and provide
informative feedback for intervention. Ambulation activities like walking, jogging, walking upstairs, walking downstairs
are performed on daily basis (Lara and Labrador, 2013). Fitness related activities are popular among the young adults
and also allows them to keep track of their fitness on a daily basis. Functional activities such as taking telephone calls,
sweeping, preparing food, taking out the trash, folding clothes, combing hair, washing hands, brushing teeth, wearing
jackets, shoes, answering the door, writing a check are the activities that every person does regularly. Inferring and
assessing such functional and behavioral activities, help decipher the personal health and wellness (Alam et al., 2016a;
Akl et al., 2015). Table 1 describes the list of existing works and the datasets, along with the specific activities and the
application areas pertaining to the proposed HAR systems.

In AR, an activity can be captured using a variety of sensors with different modalities such as video cameras,
wearable physiological and motion sensors, RADAR (Khan et al., 2016), acoustic sensors (Khan et al., 2015; Pathak et al.,
2015), Echo (Amazon Echo, 2018), everyday objects (e.g., HAPIfork (HAPIfork, 2018), food scale (SITU-The Smart Food
Nutrition Scale, 2018)), and device-free sensing (e.g., Wi-Fi (Ma et al., 2016)) etc. In addition, ambient sensors such as
infrared motion detectors and magnetic sensors have also been used extensively for AR (Cook et al., 2013). Although
video camera based HAR systems are popular for different security applications, they pose numerous challenges related
to privacy and space constraints in smart environments. For example, if the video camera is placed in a common area, like
in a corner of a room to capture the movements of a subject within its field of view, it may also capture the movements
of people who are not the subject of interest such as the caregivers or the family members. These infringes their privacy
and raises security concerns over collecting such videos. However, the most commonly used set of sensors, i.e., the
wearable sensors help eliminate the problem of privacy and security concerns for activity monitoring(Roy et al., 2013).
Wearable sensors like accelerometer and gyroscope are worn on various parts of the body, and they provide 3-axis
acceleration and orientation, respectively. Despite the fact of eliminating privacy and security concerns, the wearable
sensors also pose a set of unique challenges such as intra-class variability, inter-class similarity, class imbalance, finding
the precise start and end time of each activity (San et al., 2017) heterogeneities across the sensing devices, and device
positioning. Experiments to record data are usually conducted with multiple participants, and the data captured for
the same activity set from different participants may not be of similar nature. Therefore, the intra-class variations
become prominent. Moreover, the data pertaining to two different activities (like running and jogging) could be of similar
nature, which poses inter-class variability. Class imbalance may occur when an activity is being performed for a longer
duration than others, for example, a walking activity may be performed by the participant for a longer duration than a
jogging activity. It is also difficult to find the precise point of start and end time of an activity episode given that the
sensors usually have higher sampling frequency (hundreds of samples per second). Despite these challenges, wearable
sensors are used in the majority of the studies and researcher have been designing the appropriate methodologies and
experiments to mitigate these issues. In this study, we discuss the wearable sensors based HAR systems in the context
of new and old machine learning and data mining methodologies needed to solve some of the underpinning challenges
as mentioned previously.
RAMASAMY RAMAMURTHY ET AL . 3

HAR remains as one of the most challenging domain for the researcher owing to the complexity involved in
recognition of activities and the number of inhabitants present. Initial research on HAR has considered HAR to be a
conventional pattern recognition problem (Wang et al., 2017a). Traditional techniques like SVM, Hidden Markov models
have been extensively used in the activity recognition systems, however, there is a recent shift in the use of machine
learning and data mining techniques since the popularity of deep learning. The traditional methods (shallow learning)
requires feature engineering from the data, which is heuristic driven and heavily dependent on human knowledge of
the domain (Yang et al., 2015). This restricts the model developed for one domain to extend to another. In addition,
it is also suitable for recognizing low-level activities such as activities of daily livings, however, it is nearly impossible
to capture complex movements which involves sequence of several micro activities using shallow learning (Faridee
et al., 2018; Yang, 2009). However, deep learning methods learns the features directly from the data hierarchically
which eliminated the problem of hand-crafted feature approximations. In addition, deep learning such as Convolutional
Neural Networks have been successful in learning complex activities due to its properties of local dependencies and
scale-invariance which is elaborated by Wang et al. (2017a). HAR poses critical challenges associated with annotation
of the ground truth, recognition of activity in presence of multiple users, heterogeneity of sensing devices, faulty sensor
values and redeploying the activity model from one domain to another. Traditionally, it is required to feed the activity
model with a huge set of labeled data using a supervised machine learning algorithm so that it can learn the hidden
patterns during the training phase. Nevertheless, labeling the ground truth for a sensor data is cumbersome and always
not feasible (Hossain et al., 2016). One of the ways to tackle this challenge is to use active learning, where the model can
actively query the user for labels (Settles, 2010). Moreover, processing huge datasets incurs high computational costs
and increases the training time of the model (Cook et al., 2013). To resolve this issue, researcher started investigating
transfer learning, which allows transferring the knowledge learned from one model to another. In effect, transferring
knowledge from one model to another allows the new model to train with less amount of training samples, and hence
reduces the computational costs as well.
In this paper, we provide a comprehensive review of recent machine learning algorithms in activity recognition such
as deep learning, transfer learning, and active learning. We discuss the state-of-the-art techniques and investigate the
gaps that can help guide the future research directions. This paper is organized as follows. Section 2 discusses about
Transfer Learning in AR, followed by Active Learning in AR in Section 3. Section 4 discusses about the advances in Deep
Leaning in AR, Section 5 discusses about semantics-based techniques in AR, section 6 compares the above techniques
followed by future research directions and conclusion in Section 7 and 8, respectively.

Dataset/Literature Activities Applications

USC-HAD (Zhang and Sawchuk, 2012) Walking (forward, left, right, upstairs, down- Smart home activity recog-
stairs), running forward, jumping, sitting, nition
standing, sleeping, elevator up, elevator
down
MHealth (Banos et al., 2014) Standing still, sitting and relaxing, lying down, Smart home activity recog-
walking, climbing stairs, waist bends forward, nition
frontal elevation of arms, knees bending
(crouching), cycling, jogging, running, jump
front & back
Continued on next page
4 RAMASAMY RAMAMURTHY ET AL .

Table 1 – continued from previous page

Dataset/Literature Activities Applications
OPPORTUNITY (Roggen et al., 2009) Lying on the deckchair, get up, groom, re- Smart home activity recog-
lax, prepare coffee, drink coffee, prepare nition, functional and be-
sandwich, eat sandwich, cleanup, lie on the havioral health assessment
deckchair
PAMAP2 (Reiss and Stricker, 2012) Lying, sitting, standing, walking, running, cy- Smart home activity recog-
cling, nordic walking, watching TV, computer nition
work, car driving, ascending stairs, descend-
ing stairs, vacuum cleaning, ironing, folding
laundry, house cleaning, playing soccer, rope
jumping
Daily And Sports (Barshan and Yüksek, 2013) Sitting, standing, lying on back and on right Sports analytics
side, ascending and descending stairs, stand-
ing in an elevator still, moving around in an el-
evator, walking in a parking lot, walking on a
treadmill with a speed of 4 km/h (in flat and
15 deg inclined positions), running on a tread-
mill with a speed of 8 km/h, exercising on a
stepper, exercising on a cross trainer, cycling
on an exercise bike in horizontal and vertical
positions, rowing, jumping, playing basketball
HAR (Anguita et al., 2013) Walking (straight, upstairs, downstairs), sit- Smart home activity recog-
ting, standing, lying nition
Bicycle repair (Ogris et al., 2005) Pumping wheel, screw/unscrew, turn ped- Maintenance activity
als, turn pedals and apply break, assem- tracking, billing automa-
bling/disassembling wheel/seat, test bell, test tion, maintenance record
light generator, turn wheel, take/place item documentation
from/on carrier
Car maintenance (Stiefmeier et al., 2008) Open/close trunk, engine hood, check fuel Maintenance activity
filter cap, and other car maintenance proce- tracking, billing automa-
dures tion, maintenance record
documentation
WISDM (Kwapisz et al., 2011) Walking (Front, Upstairs, Downstairs), jog- Smart home activity recog-
ging, sitting, standing nition
Smart phone accelerometer (Do et al., 2012) Walking, running, driving, stay still Smart activity detection
for home automation,
fitness tracking
Kyoto1 (Bagaveyev and Cook, 2014) Make a phone call, Wash hands, Cook, Eat, Smart home activity recog-
Clean nition, functional and be-
havioral health assessment
Daphnet Gait (Bachlin et al., 2009) ADLs Smart home activity recog-
nition
Continued on next page
RAMASAMY RAMAMURTHY ET AL . 5

Table 1 – continued from previous page

Dataset/Literature Activities Applications
Skoda (Zappi et al., 2008) 10 manipulative gestures performed in a car Maintenance activity
maintenance scenario tracking, billing automa-
tion, maintenance record
documentation
Transportation & Physical (Bhattacharya and Walking, running, standing) and a transporta- Smart home activity
Lane, 2016) tion mode (motorized) recognition, smart home
automation
Indoor/Outdoor (Radu et al., 2014) Daily activities indoors and outdoors Smart home automa-
tion, smart home activity
recognition
Exercise Activity (Cheng et al., 2013) Bench dips, squat upright row, Dumbbells Fitness tracking, Food sug-
(DB) side raises, DB shoulder press, DB curl, gestion based on exercise
triceps extension, chest press, push up, DB fly,
bent-over row
Daily-Life Activities (Huynh et al., 2008; Sti- Sitting, standing, walking, posture upright, Smart home activity recog-
kic et al., 2011) posture kneeling, hands on table, hand above nition
chest, wrist movement, arm pendulum swing,
translation motion , cyclic motion, intense
motion, washing related, meal related
HappyFeet (Faridee et al., 2018) 10 different dance micro steps involving sub- Activity monitoring and
tle movements of limbs tracking in teaching envi-
ronments
De et al. (2015) Walk indoor, run indoor, use refrigerator, Remote monitoring of pa-
clean utensil, cooking, sit and eat, use bath- tients (Alzheimer’s disease,
room sink, indoor to outdoor, outdoor to in- bulimia, or anorexia)
door, walk (upstairs, downstairs), just stand,
lying on bed, Sit on bed, lying on floor, sit on
floor, lying on sofa, sit on sofa, sit on toilet
Dernbach et al. (2012) Cleaning, cooking, medication, sweeping, Smart home activity recog-
washing hands, watering plants nition
Pham and Olivier (2009) Chopping, peeling, slicing, dicing, coring, Smart home activity recog-
spreading, eating, stirring, scooping, scraping nition, kitchen activity
and shaving recognition
CASAS (Singla et al., 2009) Fill medication dispenser, watch DVD, wa- Smart home activity recog-
ter plants, converse on phone, write birthday nition, functional and be-
card, prepare meal, sweep and dust, select an havioral health assessment
outfit
SmartFaber (Riboni et al., 2016) Taking medicines, cooking, eating Smart home activity
recognition, Cognitive
assessment
TA B L E 1 Activitles, applications and associated literature/datasets
6 RAMASAMY RAMAMURTHY ET AL .

FIGURE 1 Transfer Learning Illustration

FIGURE 2 Effect of Transfer learning

2 | TRANSFER LEARNING IN AR

Transfer learning can be defined as the ability to extend what has been learned in one context to new contexts (Byrnes, 2001).
Woodworth and Thorndike (1901) first explored how individuals transfer learned concepts between different contexts that share
common features. Barnett and Ceci (2002) provided a taxonomy of features that influences transfer learning in humans. In the
field of machine learning, transfer learning is interchangeably used with different names like learning to learn, life-long learning,
knowledge transfer, inductive transfer, context-sensitive learning, and meta-learning (Cook et al., 2013).

Transfer learning lets us transfer knowledge from one domain to another assuming that there exists some relationship be-
tween the source and target areas which allows for the successful transfer of knowledge from the source to the target. Figure
1 illustrates three different scenarios where transfer learning can be applied. In scenario-1, the activity is cycling, however, the
gender of the person performing the activity is different. The data acquired for the same activity is different when performed by
different persons. In scenario-2, the device that is used to capture the activity data is different, i.e., smart phone and smart watch.
In Scenario-3, the ambiance where the activity is performed is different, indoors and outdoors. The information extracted by a
model in the source domain can help train the model in the target domain with less amount of annotated data. Figure 2 repre-
sents the effect of transfer learning, where a reduced amount of data points might be required to train the model in target domain
because of using information from the previously trained model. This dramatically reduces the computational costs and the anno-
tating efforts. The various scenarios where transfer learning can be applied is explained comprehensively by Cook et al. (2013).
RAMASAMY RAMAMURTHY ET AL . 7

The authors have elaborated on the different modalities, data labeling process and the taxonomy of type of knowledge trans-
ferred in transfer learning based AR. Khan and Roy (2017) have investigated transferring knowledge among the models having
different probability distributions. The authors have evaluated their AR models using random forest, decision tree and transfer
boost algorithms and have used accuracy as an evaluation metric to assess the performance of the model. The authors tested this
methodology on HAR (Anguita et al., 2013), Daily And Sports (Barshan and Yüksek, 2013), MHealth (Banos et al., 2014) datasets
and have proved that a HAR model can be trained using a reduced number of instances. The probability distribution of the ac-
celerometer data varies heavily among different users, and the performance of a model will degrade if the model is trained for a
person and tested on another. In an attempt to address this problem, Deng et al. (2014) have proposed a cross-person activity
recognition model that integrates transfer learning and Reduced Kernel Extreme Learning Machine (RKELM). RKELM is popular
and is effective when the dataset is extremely huge. It randomly selects a subset of the dataset and analytically computes the
weights of the classifier, therefore reduces the computational time and also provides a good approximation of the AR model to
that generated from actual data. The methodology was assessed on Activities of Daily Living (ADL) dataset. The authors, Wang
et al. (2017b), have introduced a new framework to transfer the labeled activity data from source domain to target domain. The
model, Stratified Transfer Learning transforms source and target domain into the same subspace where the data distribution is com-
parable followed by cross-domain activity recognition and assessed on OPPORTUNITY (Roggen et al., 2009), PAMAP2 (Reiss and
Stricker, 2012), and Daily And Sports (Barshan and Yüksek, 2013) datasets.
A scenario where the deployment context may be different from the learning context is inevitable. An illustration of such
scenario could be when the jogging activity is performed on a treadmill and the same on the streets. The former is performed in
a controlled setting where the speed is adjusted to a constant speed, and there may not be any obstructions, however, the latter
could suffer from many obstructions and variable speed as they cannot be controlled by us. Diethe et al. (2016) have proposed
a hierarchical Bayesian transfer learning model and have also addressed the problem of accurately labeling the data using active
learning. The authors have evaluated the model using HAR using smart phone dataset (Anguita et al., 2013) as their source and
USC-HAD dataset (Zhang and Sawchuk, 2012) as target. Ying et al. (2015) have proposed a transfer learning model on high variety
data (data from different sources) which is validated using statistical hypothesis Kolmogorov-Smirnov and χ 2 goodness of fit test,
and evaluated on Walk8, HAR (Anguita et al., 2013), and DaSA (Altun et al., 2010) datasets. Rokni and Ghasemzadeh (2017) have
proposed an approach for autonomous retraining of machine learning algorithms without any new labeled training data. The new
data considered in this study is from another sensor that is added to the system, and the machine learning algorithms used in this
study are Decision Trees, k-Nearest Neighborhood, and SVM. The model was evaluated on OPPORTUNITY (Roggen et al., 2009)
dataset and DaSA (Altun et al., 2010) dataset.
AR is highly dependent on numerous factors such as the type of sensor used, the environmental setting, the experimental
settings and so on. If a model is trained using one combination of the above settings, that model can be evaluated for a different
combination of settings using transfer learning. Transfer learning is still an emerging topic, and researchers are still exploiting the
field to discover the applicability of using it for large scale cross-domain human activity recognition.

3 | ACTIVE LEARNING IN AR

Active learning in AR is a recently emerging field. The active learning algorithms aim at mitigating the learning complexity and cost.
It helps to select an optimal number of informative unlabeled data samples and query the annotator for the labels. It minimizes
labeling effort and elevates the prediction accuracy (Hossain et al., 2016). Figure 3 shows an illustration of an active learning
enabled model. Active learning has been popular in other fields, however in AR; only a few researchers have been working on
active learning. Alemdar et al. (2017) has investigated three different techniques (least confidence method, margin sampling, and
entropy-based) to find the most informative unlabeled data samples and have proved that the annotation effort has been reduced
by a factor of 2-4 times. Hossain et al. (2016) have proposed a dynamic k-means clustering algorithm based active learning ap-
proach, and have used uncertainty sampling to find the most informative unlabeled data samples, and validated the proposed
approach using real-life data traces. The authors have also proposed a data representation technique for crowd-sourcing the la-
beling, and they have discussed its repercussion on active learning. Bannach et al. (2017) have investigated the self-adaptability
8 RAMASAMY RAMAMURTHY ET AL .

FIGURE 3 Active Learning Cycle

of AR model when a new sensor is introduced to the system and have evaluated their model using a bicycle repair (Ogris et al.,
2005), a car maintenance (Stiefmeier et al., 2008) and OPPORTUNITY (Roggen et al., 2009) datasets. Abdallah et al. (2015) have
proposed a personalized and adaptive framework for AR that incrementally helps learn the activity model from high speed, multi-
dimensional streaming data. It recognized personalized user’s activities using active learning, employed ensemble classifier to
train the model and evaluated the model using OPPORTUNITY (Roggen et al., 2009), WISDM (Kwapisz et al., 2011), and smart-
phone accelerometer datasets (Do et al., 2012). Bagaveyev and Cook (2014) have investigated two different approaches to select
unlabeled data for annotations using Expected Entropy and Query by Committee for active learning, used random forest based
classifier for activity inference, and validated the results on an in house dataset, Kyoto1. Active learning in AR is still emerging
field and has plenty of scope in the future.

4 | DEEP LEARNING IN AR

In HAR systems designed using shallow learning, the frequently used feature heuristics are dependent on the domain knowledge
of the researcher and the performance of the machine learning techniques is highly dependent on the data representation (Bengio,
2013). The commonly used features are time domain features (mean, variance, time sequences) (Bulling et al., 2014), frequency
domain features (Fourier transform, entropy) and other transformations (wavelet transform) (Huynh and Schiele, 2005). How-
ever, in deep learning, the features are learned from the raw data hierarchically by performing some nonlinear transformation.
The nonlinear transformation determines the type of deep learning network. Deep learning has been popular in the last few years,
and numerous work has been done using deep learning in AR. The popular deep learning techniques in AR include Deep Neural
Network (DNN), Convolutional Neural Networks(CNN), Recurrent Neural Networks(RNN), Long- Short-Term Memory (LSTM)
RNN networks. Wang et al. (2017a) has comprehensively described the deep learning techniques and the effect of applying deep
learning to time series activity signals. Hammerla et al. (2016) have explored DNN, CNN, and RNN for activity dataset and con-
cluded that the RNN performed better than the state-of-the-art results for OPPORTUNITY (Roggen et al., 2009), PAMAP2 (Reiss
and Stricker, 2012) and Daphnet Gait (Bachlin et al., 2009) dataset. The authors also found that the RNN outperformed CNN for
activities which are of short duration. Ordóñez and Roggen (2016) have proposed a novel deep network comprising of convolu-
tional and LSTM layers. The authors optimized the hyper-parameters of their network and fused the various sensor modalities
like accelerometer, gyroscope, magnetometer in different combinations and compared the results by evaluating them on OPPOR-
TUNITY (Roggen et al., 2009) and Skoda (Zappi et al., 2008) datasets. Ronao and Cho (2016) have proposed multilayer CNN
model with alternating convolutional and pooling layers and showed that their proposed model outperforms the state-of-the-art
accuracy for ADLs which were recorded by the authors from 30 different users. Ravi et al. (2016) have used Short Term Fourier
Transform of the accelerometer data as an input to the proposed CNN network and have achieved accuracy close to the state-
of-the-art results. Bhattacharya and Lane (2016) have designed and developed a Restricted Boltzmann Machine (RBM) based
RAMASAMY RAMAMURTHY ET AL . 9

activity recognition model for smart watches and have proved that the model does not have any hardware constraints. The au-
thors have also validated their accuracy with real-life data sets such as OPPORTUNITY (Roggen et al., 2009), Transportation &
Physical (Bhattacharya and Lane, 2016), Indoor/Outdoor (Radu et al., 2014) datasets with different state-of-the-art classifiers for
each of the activities.

The most recent contributions towards deep learning (Panwar et al., 2017) postulated a novel technique of using ensembles
of deep LSTM networks using wearable sensing data. This is the only work using an ensemble of deep learning techniques so it
has scope of further research in this direction. Sani et al. (2017) have evaluated the hand-crafted features and the CNN derived
features using kNN using activities of daily living sensor data collected from the wrist and the waist. Studying further about
the features derived from a deep learning network and comparing them with the heuristic features will help the researcher to
understand the deep learning networks better. San et al. (2017) have proposed a multichannel CNN architecture for multiple
sensor data. The authors have evaluated their results with Decision Tree, kNN, and Naive Bayes classifiers and have seen drastic
improvements in terms of activity recognition accuracy. Ensemble deep learning models, LSTM, and RNN have not been well
investigated for multichannel multi-modal sensor data and therefore, this would be a future direction for research.

5 | SEMANTICS IN AR

Sensor-based activity recognition approaches can be classified as data-driven and knowledge-driven approaches. A data-driven
approach involves data collection, followed by extracting knowledge from the data by performing techniques like basic statistical
techniques or other data mining techniques. On the other hand, knowledge-driven approaches use prior domain knowledge fol-
lowed by the application specific knowledge on the sensory data. One such commonly used model is Ontology modeling (Liu et al.,
2016). One of the applications of knowledge-driven activity recognition is in smart home settings. In a home, any person tends to
perform a lot of activities pertaining to time, location, context and so on. For example, a person brushes the teeth in the morning,
and at night in the bathroom, a person cooks food in the kitchen. The domain knowledge in these cases is the location and the tim-
ing of the activity performed. This domain knowledge allows the AR system to correctly detect the activities. (Riboni et al., 2016)
has proposed a unsupervised approach to recognize complex activities by exploiting the semantic relationship between activities
and smart-home environment, context data and sensing devices. The authors tested the model on CASAS (Singla et al., 2009) and
SmartFaber (Riboni et al., 2016) datasets, and achieved an accuracy that is comparable to the supervised state-of-the-art algo-
rithm. (Gayathri et al., 2017) leveraged the strengths of ontological modeling through Markov Logic Network and its probabilistic
reasoning and tested the model on CASAS (Singla et al., 2009) dataset. (Villalonga et al., 2017) investigated ontology-based sensor
selection for real world AR that can be used to select the best sensor to capture the activity better. (Woznowski et al., 2016) have
presented a hierarchical ontological modeling for annotating activity data and further discussed the labeling strategies and the
best practices. Ye et al. (2015) have proposed an algorithm that constructs domain ontologies profiles and, extract the semantic
features form sensor events into spatial, temporal and thematic aspects. The activities are then recognized by matching seg-
mented sensor sequences to ontological profiles. The authors have evaluated using the Interleaved ADL Activities (IAA) dataset
(Cook and Schmitter-Edgecombe, 2009) which is described in Table 1. The authors Liu et al. (2016) have presented an algorithm
to identify the complex high-level activities from simple low-level actions in the sensor domain. The algorithm computes the sup-
port after adding the subsequent patterns iteratively to build a feature space that can be fed to a SVM classifier. The proposed
technique has been evaluated on OPPORTUNITY (Roggen et al., 2009) dataset. In another study, the authors combined the data-
driven and knowledge-driven approaches which posit unsupervised techniques to discover sequential activity patterns based on
the learned ontological model. Cheng et al. (2013) proposed a zero-shot learning framework based on semantic sequences of
data that considers both hierarchical and sequential nature of the activity to detect unseen activities. The model was evaluated
on their own Exercise Activity Dataset, and Daily-Life Activities dataset. De et al. (2015) have proposed a conditional random
field classifier for activity prediction which captures temporal relationships in activity time series that helps to capture complex
activities. The technique was evaluated using an in-house dataset comprising of 19 activities listed in Table 1. Since knowledge-
driven approaches combine information from various sources to build the domain knowledge, it can be used for multi-inhabitant
activity recognition.
10 RAMASAMY RAMAMURTHY ET AL .

One of the major challenges in multi-inhabitant activity recognition in a smart home environment is that the ambient sen-
sors are susceptible to recording data pertaining to non-subjects. Despite the disadvantages, there has been some research on
the same. The authors Roy et al. (2013) have combined the ambient sensors along with the body-worn sensors to extract person-
independent context and person-specific context of activity and used Hidden Markov Model to detect the activities. The ambi-
ent sensors detect the movements of the subject, however, in addition, it also captures the movements of others who live in its
range in a multi-inhabitant environment. Alam et al. (2016b) proposed a probabilistic Hierarchical Dynamic Bayesian Network
(HDBN) to combine the postural and gestural micro-activities, and further extended to a multi-inhabitant framework using cou-
pled - HDBN. The authors further discovered the spatiotemporal constraints for activities of users in the multi-inhabitant envi-
ronment and evaluated their model with a dataset collected from real-life scenarios. The multi-inhabitant activity recognition for
smart home environment is relatively a new field, and further work based on ambient and wearable sensors is required to build a
robust knowledge-driven system. Building semantic knowledge requires domain knowledge, and it has to be well represented in
such a way that the data mining techniques can understand the semantics which is still a challenging problem and requires further
studies.

6 | COMPARATIVE DISCUSSION

All the techniques discussed so far for human activity recognition have its own merits and demerits. Transfer learning allows
transferring knowledge, which addresses the problems of class imbalance, insufficient annotated data, domain adaptation, scal-
able model construction etc. However, for transfer learning to be effective, the quality of the source data is crucial. The activity
event start and end time mismatch with the annotations, motion artifacts caused by wearable sensors, intra-class variability are
few major concerns that hinders the performance of transfer learning in AR. Active learning helps ease the annotation efforts by
querying the labels of informative samples only and therefore reduce the annotation efforts. However, active learning requires
a high quality data similar to transfer learning. In addition, selection of an appropriate criteria for selection of informative data
points is challenging. The error in re-annotating informative data points leads to propagation of error to the model that tries to
learn the activity. Active learning poses another challenge of selection of annotators. Different annotators may label the sequence
of data differently depending on their expertise therefore there is a need for annotator selection model (Hossain et al., 2016).
Another most recent trend in AR is deep learning. It is noted that deep learning has been outperforming traditional machine
learning methods as deep learning is able to extract the features from the raw data in contrast to expensive feature engineering
in traditional techniques. Feature engineering requires domain knowledge, which leads to approximations of the features and
makes them sensitive to the challenges related to AR as discussed in Section 1 in comparison to deep learning where the features
are hierarchically learned directly from the raw data. Despite the advantages, deep learning methods has high computational
requirements. Deep learning for mobile and low powered devices is an emerging field and it is also a potential area for future
research. In all of the above techniques, the objective was to learn/classify/detect the activities. However, the context and the
intention of doing the activity was not discussed. Semantic based techniques utilize the prior knowledge and integrate with the
sensory data to infer the context and the intention of performing the activity. This knowledge-driven approach allows HAR sys-
tems to be used in a variety of applications as discussed in Section 1. Nevertheless, these approaches are sensitive to the prior
knowledge as the context is derived from it.

7 | FUTURE DIRECTIONS

The nature of the activity data is such that they are self-similar fractal patterns and they repeat over time. The authors of Gupta
and Dallas (2014) have used detrended fluctuation analysis coefficients as features in learning their classifier and Sekine et al.
(2002) have used the fractal dimensions of wavelet coefficients to differentiate three different walking styles, age groups, and
patients suffering from Parkinson disease. As fractal analysis has been explored and proven successful in analyzing physiological
data like heart rate variability, and have shown promising results in Gupta and Dallas (2014); Sekine et al. (2002), it is a direction
RAMASAMY RAMAMURTHY ET AL . 11

that needs further investigation for activity recognition.

Another future direction in activity recognition could be synthetic data generation. In most cases, for the data mining tech-
nique to learn the model effectively, it is essential to train the model with a huge dataset. In addition, as discussed in Section 1, the
data could suffer from challenges like intra-class variability and class imbalance. This occurs due to the fact that some activities
take a longer time when compared to the others. In a real scenario, a person may walk for a longer duration when compared to
jogging. The duration varies drastically for activities like sweeping, cleaning the house, cooking, bathing, jogging, walking upstairs
and downstairs, and so on. Generating synthetic data could help solve some of the above problems. Few researcher have already
worked on generating synthetic data for activity recognition. Mendez-Vazquez et al. (2009) have used Markov chains to generate
the patterns of activities and Poisson processes to generate the time stamps. This study is extensive and considers activities such
as walking, reading, sleep, sitting, housework, and exercise. Activity data shows subtle variations, even among the data collected
from the same subject for the same activity at a different time, for which generating data synthetically might be a challenging prob-
lem. Alzantot et al. (2017) have proposed a synthetic data generation model using deep learning for ADLs using smartphones. The
authors have presented a LSTM and Mixture Density Network based generator model to generate the synthetic data and a LSTM
based discriminative model that distinguishes between true and the synthesized data. Synthetic data generation requires further
investigation to enumerate data patterns that give a better representation of the raw signals for a variety of activity signals so
that the AR models can extract maximum information from the data.
Another possible area for HAR could be in processing and fusing data from heterogeneous devices. In recent times, there are
a huge number of off-the-shelf devices are available such as Actigraph (ActiGraph, 2018), Microsoft Band (Microsoft Band, 2018),
Empatica E4 (Empatica E4, 2018), Fitbit (Fitbit, 2018), Google Home (Google Home, 2018), Amazon Echo (Amazon Echo, 2018).
Often it is the case that a model that is developed using the data collected using one device does not perform well with the other
devices. This occurs due to the heterogeneities among the devices. The heterogeneities among various mobile devices, smart
watches have been discussed in detail by Stisen et al. (2015). The authors have investigated a large number of devices to study
device heterogeneities and have proposed techniques to mitigate them. The heterogeneities they address include sampling rate
heterogeneity, sampling rate instability and sensor biases. Investigating cross-domain transfer and deep learning to accommodate
heterogeneity in the learned models are potential future directions in activity recognition (Khan and Roy, 2018; Khan et al., 2018).

8 | CONCLUSION

Mining the activity data to detect and understand the activity is of utmost importance as activity recognition finds its applica-
tion in various fields such as personal healthcare like fitness tracking, fall detection of elderly people, monitoring functional and
behavioral health using wearables. In this paper, we articulate the recent trends in AR towards addressing the limitations of the
traditional machine learning algorithms and mitigating few system design challenges. We note that deep learning architectures
have been used largely due to its advantage of hierarchically self-derived features, which help represent the data better compared
to the handcrafted features. Therefore, it is highly important to design and develop robust data mining techniques to extract the
knowledge and machine learning techniques to infer and validate that knowledge from data which will allow the AR system to
make intelligent decisions. This study presents the recent trends and developments in machine learning techniques, to address
the next-generation activity recognition challenges across many devices, systems, persons and environments.

REFERENCES

Abdallah, Z. S., Gaber, M. M., Srinivasan, B. and Krishnaswamy, S. (2015) Adaptive mobile activity recognition system with
evolving data streams. Neurocomputing, 150, 304–317.

ActiGraph (2018) URL: http://www.actigraphcorp.com/.

Akl, A., Taati, B. and Mihailidis, A. (2015) Autonomous unobtrusive detection of mild cognitive impairment in older adults. IEEE
Transactions on Biomedical Engineering, 62, 1383–1394.
12 RAMASAMY RAMAMURTHY ET AL .

Alam, M. A. U., Roy, N., Holmes, S., Gangopadhyay, A. and Galik, E. (2016a) Automated functional and behavioral health assess-
ment of older adults with dementia. In Connected Health: Applications, Systems and Engineering Technologies (CHASE), 2016
IEEE First International Conference on, 140–149. IEEE.

Alam, M. A. U., Roy, N., Misra, A. and Taylor, J. (2016b) Cace: Exploiting behavioral interactions for improved activity recogni-
tion in multi-inhabitant smart homes. In Distributed Computing Systems (ICDCS), 2016 IEEE 36th International Conference on,
539–548. IEEE.

Alemdar, H., van Kasteren, T. and Ersoy, C. (2017) Active learning with uncertainty sampling for large scale activity recognition
in smart homes. Journal of Ambient Intelligence and Smart Environments, 9, 209–223.

Altun, K., Barshan, B. and Tunçel, O. (2010) Comparative study on classifying human activities with miniature inertial and
magnetic sensors. Pattern Recognition, 43, 3605–3620.

Alzantot, M., Chakraborty, S. and Srivastava, M. (2017) Sensegen: A deep learning architecture for synthetic sensor data gen-
eration. In Pervasive Computing and Communications Workshops (PerCom Workshops), 2017 IEEE International Conference on,
188–193. IEEE.

Amazon Echo (2018) URL: https://www.amazon.com/echo/.

Anguita, D., Ghio, A., Oneto, L., Parra, X. and Reyes-Ortiz, J. L. (2013) A public domain dataset for human activity recognition
using smartphones. In ESANN.

Bachlin, M., Roggen, D., Troster, G., Plotnik, M., Inbar, N., Meidan, I., Herman, T., Brozgol, M., Shaviv, E., Giladi, N. et al. (2009)
Potentials of enhanced context awareness in wearable assistants for parkinson’s disease patients with the freezing of gait
syndrome. In Wearable Computers, 2009. ISWC’09. International Symposium on, 123–130. IEEE.

Bagaveyev, S. and Cook, D. J. (2014) Designing and evaluating active learning methods for activity recognition. In Proceedings
of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication, 469–478. ACM.

Bannach, D., Jänicke, M., Rey, V. F., Tomforde, S., Sick, B. and Lukowicz, P. (2017) Self-adaptation of activity recognition systems
to new sensors. arXiv preprint arXiv:1701.08528.

Banos, O., Garcia, R., Holgado-Terriza, J. A., Damas, M., Pomares, H., Rojas, I., Saez, A. and Villalonga, C. (2014) mhealthdroid: a
novel framework for agile development of mobile health applications. In International Workshop on Ambient Assisted Living,
91–98. Springer.

Barnett, S. M. and Ceci, S. J. (2002) When and where do we apply what we learn?: A taxonomy for far transfer. Psychological
bulletin, 128, 612.

Barshan, B. and Yüksek, M. C. (2013) Recognizing daily and sports activities in two open source machine learning environments
using body-worn sensor units. The Computer Journal, 57, 1649–1667.

Bengio, Y. (2013) Deep learning of representations: Looking forward. In International Conference on Statistical Language and
Speech Processing, 1–37. Springer.

Bhattacharya, S. and Lane, N. D. (2016) From smart to deep: Robust activity recognition on smartwatches using deep learning.
In Pervasive Computing and Communication Workshops (PerCom Workshops), 2016 IEEE International Conference on, 1–6. IEEE.

Bulling, A., Blanke, U. and Schiele, B. (2014) A tutorial on human activity recognition using body-worn inertial sensors. ACM
Computing Surveys (CSUR), 46, 33.

Byrnes, J. P. (2001) Cognitive development and learning in instructional contexts. Allyn & Bacon.

Cheng, H.-T., Sun, F.-T., Griss, M., Davis, P., Li, J. and You, D. (2013) Nuactiv: Recognizing unseen new activities using semantic
attribute-based learning. In Proceeding of the 11th annual international conference on Mobile systems, applications, and services,
361–374. ACM.
RAMASAMY RAMAMURTHY ET AL . 13

Cook, D., Feuz, K. D. and Krishnan, N. C. (2013) Transfer learning for activity recognition: A survey. Knowledge and information
systems, 36, 537–556.

Cook, D. J. and Schmitter-Edgecombe, M. (2009) Assessing the quality of activities in a smart environment. Methods of infor-
mation in medicine, 48, 480.

De, D., Bharti, P., Das, S. K. and Chellappan, S. (2015) Multimodal wearable sensing for fine-grained activity recognition in
healthcare. IEEE Internet Computing, 19, 26–35.

Deng, W.-Y., Zheng, Q.-H. and Wang, Z.-M. (2014) Cross-person activity recognition using reduced kernel extreme learning
machine. Neural Networks, 53, 1–7.

Dernbach, S., Das, B., Krishnan, N. C., Thomas, B. L. and Cook, D. J. (2012) Simple and complex activity recognition through
smart phones. In Intelligent Environments (IE), 2012 8th International Conference on, 214–221. IEEE.

Diethe, T., Twomey, N. and Flach, P. (2016) Active transfer learning for activity recognition. In European Symposium on Artificial
Neural Networks, Computational Intelligence and Machine Learning.

Do, T. M., Loke, S. W. and Liu, F. (2012) Healthylife: An activity recognition system with smartphone using logic-based stream
reasoning. In International Conference on Mobile and Ubiquitous Systems: Computing, Networking, and Services, 188–199.
Springer.

Empatica E4 (2018) URL: https://www.empatica.com/en-eu/research/e4/.

Faridee, A. Z. M., Ramasamy Ramamurthy, S., Hossain, H. M. S. and Roy, N. (2018) Happyfeet: Recognizing and assessing dance
on the floor. In HotMobile’18: 19th International Workshop on Mobile Computing Systems & Applications, February 12–13,
2018, Tempe , AZ, USA. ACM.

Fitbit (2018) URL: https://www.fitbit.com/home.

Gayathri, K., Easwarakumar, K. and Elias, S. (2017) Probabilistic ontology based activity recognition in smart homes using
markov logic network. Knowledge-Based Systems, 121, 173–184.

Google Home (2018) URL: https://store.google.com/product/google_home.

Gupta, P. and Dallas, T. (2014) Feature selection and activity recognition system using a single triaxial accelerometer. IEEE
Transactions on Biomedical Engineering, 61, 1780–1786.

Hammerla, N. Y., Halloran, S. and Ploetz, T. (2016) Deep, convolutional, and recurrent models for human activity recognition
using wearables. arXiv preprint arXiv:1604.08880.

HAPIfork (2018) URL: https://www.hapi.com/product/hapifork.

Hossain, H. S., Khan, M. A. A. H. and Roy, N. (2016) Active learning enabled activity recognition. Pervasive and Mobile Computing.

Huynh, T., Fritz, M. and Schiele, B. (2008) Discovery of activity patterns using topic models. In Proceedings of the 10th interna-
tional conference on Ubiquitous computing, 10–19. ACM.

Huynh, T. and Schiele, B. (2005) Analyzing features for activity recognition. In Proceedings of the 2005 joint conference on Smart
objects and ambient intelligence: innovative context-aware services: usages and technologies, 159–163. ACM.

Khan, M. A. A. H., Hossain, H. and Roy, N. (2015) Infrastructure-less occupancy detection and semantic localization in smart
environments. In proceedings of the 12th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Network-
ing and Services on 12th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services,
51–60. ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering).
14 RAMASAMY RAMAMURTHY ET AL .

Khan, M. A. A. H., Kukkapalli, R., Waradpande, P., Kulandaivel, S., Banerjee, N., Roy, N. and Robucci, R. (2016) Ram: Radar-based
activity monitor. In INFOCOM 2016-The 35th Annual IEEE International Conference on Computer Communications, IEEE, 1–9.
IEEE.

Khan, M. A. A. H. and Roy, N. (2017) Transact: Transfer learning enabled activity recognition. In Pervasive Computing and Com-
munications Workshops (PerCom Workshops), 2017 IEEE International Conference on, 545–550. IEEE.

— (2018) Untran: Recognizing unseen activities with unlabeled data using transfer learning. In International Conference on
Internet-of-Things Design and Implementation (IoTDI). ACM/IEEE.

Khan, M. A. A. H., Roy, N. and Misra, A. (2018) Scaling human activity recognition via deep learning-based domain adaptation.
In In Proceedings of the IEEE International Conference on Pervasive Computing and Communications (PerCom). IEEE.

Kwapisz, J. R., Weiss, G. M. and Moore, S. A. (2011) Activity recognition using cell phone accelerometers. ACM SigKDD Explo-
rations Newsletter, 12, 74–82.

Lara, O. D. and Labrador, M. A. (2013) A survey on human activity recognition using wearable sensors. IEEE Communications
Surveys and Tutorials, 15, 1192–1209.

Liu, Y., Nie, L., Liu, L. and Rosenblum, D. S. (2016) From action to activity: Sensor-based activity recognition. Neurocomputing,
181, 108–115.

Ma, J., Wang, H., Zhang, D., Wang, Y. and Wang, Y. (2016) A survey on wi-fi based contactless activity recognition. In Ubiquitous
Intelligence & Computing, Advanced and Trusted Computing, Scalable Computing and Communications, Cloud and Big Data Com-
puting, Internet of People, and Smart World Congress (UIC/ATC/ScalCom/CBDCom/IoP/SmartWorld), 2016 Intl IEEE Conferences,
1086–1091. IEEE.

Mendez-Vazquez, A., Helal, A. and Cook, D. (2009) Simulating events to generate synthetic data for pervasive spaces. In Work-
shop on Developing Shared Home Behavior Datasets to Advance HCI and Ubiquitous Computing Research.

Microsoft Band (2018) URL: https://www.microsoft.com/en-us/band.

Ogris, G., Stiefmeier, T., Junker, H., Lukowicz, P. and Troster, G. (2005) Using ultrasonic hand tracking to augment motion anal-
ysis based recognition of manipulative gestures. In Wearable Computers, 2005. Proceedings. Ninth IEEE International Sympo-
sium on, 152–159. IEEE.

Ordóñez, F. J. and Roggen, D. (2016) Deep convolutional and lstm recurrent neural networks for multimodal wearable activity
recognition. Sensors, 16, 115.

Panwar, M., Dyuthi, S. R., Prakash, K. C., Biswas, D., Acharyya, A., Maharatna, K., Gautam, A. and Naik, G. R. (2017) Cnn based
approach for activity recognition using a wrist-worn accelerometer. In Engineering in Medicine and Biology Society (EMBC),
2017 39th Annual International Conference of the IEEE, 2438–2441. IEEE.

Pathak, N., Khan, M. A. A. H. and Roy, N. (2015) Acoustic based appliance state identifications for fine-grained energy analytics.
In Pervasive Computing and Communications (PerCom), 2015 IEEE International Conference on, 63–70. IEEE.

Pham, C. and Olivier, P. (2009) Slice&dice: Recognizing food preparation activities using embedded accelerometers. In Euro-
pean Conference on Ambient Intelligence, 34–43. Springer.

Radu, V., Katsikouli, P., Sarkar, R. and Marina, M. K. (2014) A semi-supervised learning approach for robust indoor-outdoor
detection with smartphones. In Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems, 280–294.
ACM.

Ravi, D., Wong, C., Lo, B. and Yang, G.-Z. (2016) Deep learning for human activity recognition: A resource efficient implementa-
tion on low-power devices. In Wearable and Implantable Body Sensor Networks (BSN), 2016 IEEE 13th International Conference
on, 71–76. IEEE.
RAMASAMY RAMAMURTHY ET AL . 15

Reiss, A. and Stricker, D. (2012) Introducing a new benchmarked dataset for activity monitoring. In Wearable Computers (ISWC),
2012 16th International Symposium on, 108–109. IEEE.

Riboni, D., Sztyler, T., Civitarese, G. and Stuckenschmidt, H. (2016) Unsupervised recognition of interleaved activities of daily
living through ontological and probabilistic reasoning. In Proceedings of the 2016 ACM International Joint Conference on
Pervasive and Ubiquitous Computing, 1–12. ACM.

Roggen, D., Forster, K., Calatroni, A., Holleczek, T., Fang, Y., Troster, G., Ferscha, A., Holzmann, C., Riener, A., Lukowicz, P. et al.
(2009) Opportunity: Towards opportunistic activity and context recognition systems. In World of Wireless, Mobile and Mul-
timedia Networks & Workshops, 2009. WoWMoM 2009. IEEE International Symposium on a, 1–6. IEEE.

Rokni, S. A. and Ghasemzadeh, H. (2017) Synchronous dynamic view learning: a framework for autonomous training of activity
recognition models using wearable sensors. In IPSN, 79–90.

Ronao, C. A. and Cho, S.-B. (2016) Human activity recognition with smartphone sensors using deep learning neural networks.
Expert Systems with Applications, 59, 235–244.

Roy, N., Misra, A. and Cook, D. (2013) Infrastructure-assisted smartphone-based adl recognition in multi-inhabitant smart
environments. In Pervasive Computing and Communications (PerCom), 2013 IEEE International Conference on, 38–46. IEEE.

San, P. P., Kakar, P., Li, X.-L., Krishnaswamy, S., Yang, J.-B. and Nguyen, M. N. (2017) Deep learning for human activity recognition.

Sani, S., Wiratunga, N. and Massie, S. (2017) Learning deep features for knn-based human activity recognition.

Sekine, M., Tamura, T., Akay, M., Fujimoto, T., Togawa, T. and Fukui, Y. (2002) Discrimination of walking patterns using wavelet-
based fractal analysis. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 10, 188–196.

Settles, B. (2010) Active learning literature survey. University of Wisconsin, Madison, 52, 11.

Singla, G., Cook, D. J. and Schmitter-Edgecombe, M. (2009) Tracking activities in complex settings using smart environment
technologies. International journal of biosciences, psychiatry, and technology (IJBSPT), 1, 25.

SITU-The Smart Food Nutrition Scale (2018) URL: http://situscale.com/.

Stiefmeier, T., Roggen, D., Ogris, G., Lukowicz, P. and Tröster, G. (2008) Wearable activity tracking in car manufacturing. IEEE
Pervasive Computing, 7.

Stikic, M., Larlus, D., Ebert, S. and Schiele, B. (2011) Weakly supervised recognition of daily life activities with wearable sensors.
IEEE transactions on pattern analysis and machine intelligence, 33, 2521–2537.

Stisen, A., Blunck, H., Bhattacharya, S., Prentow, T. S., Kjærgaard, M. B., Dey, A., Sonne, T. and Jensen, M. M. (2015) Smart
devices are different: Assessing and mitigatingmobile sensing heterogeneities for activity recognition. In Proceedings of
the 13th ACM Conference on Embedded Networked Sensor Systems, 127–140. ACM.

Villalonga, C., Pomares, H., Rojas, I. and Banos, O. (2017) Mimu-wear: Ontology-based sensor selection for real-world wear-
able activity recognition. Neurocomputing, 250, 76–100.

Wang, J., Chen, Y., Hao, S., Peng, X. and Hu, L. (2017a) Deep learning for sensor-based activity recognition: A survey. arXiv
preprint arXiv:1707.03502.

Wang, J., Chen, Y., Hu, L., Peng, X. and Yu, P. S. (2017b) Stratified transfer learning for cross-domain activity recognition. arXiv
preprint arXiv:1801.00820.

Woodworth, R. S. and Thorndike, E. (1901) The influence of improvement in one mental function upon the efficiency of other
functions.(i). Psychological review, 8, 247.
16 RAMASAMY RAMAMURTHY ET AL .

Woznowski, P., King, R., Harwin, W. and Craddock, I. (2016) A human activity recognition framework for healthcare applica-
tions: ontology, labelling strategies, and best practice. In Proceedings of the International Conference on Internet of Things and
Big Data (IoTBD), 369–377.

Yang, J., Nguyen, M. N., San, P. P., Li, X. and Krishnaswamy, S. (2015) Deep convolutional neural networks on multichannel time
series for human activity recognition. In IJCAI, 3995–4001.

Yang, Q. (2009) Activity recognition: linking low-level sensors to high-level intelligence. In IJCAI, vol. 9, 20–25.

Ye, J., Stevenson, G. and Dobson, S. (2015) Kcar: A knowledge-driven approach for concurrent activity recognition. Pervasive
and Mobile Computing, 19, 47–70.

Ying, J. J.-C., Lin, B.-H., Tseng, V. S. and Hsieh, S.-Y. (2015) Transfer learning on high variety domains for activity recognition. In
Proceedings of the ASE BigData & SocialInformatics 2015, 37. ACM.

Zappi, P., Lombriser, C., Stiefmeier, T., Farella, E., Roggen, D., Benini, L. and Troster, G. (2008) Activity recognition from on-body
sensors: accuracy-power trade-off by dynamic sensor selection. Lecture Notes in Computer Science, 4913, 17.

Zhang, M. and Sawchuk, A. A. (2012) Usc-had: a daily activity dataset for ubiquitous activity recognition using wearable sen-
sors. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing, 1036–1043. ACM.