Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
7 views65 pages

Dissertation Updated Copy - 6607133

This MSc dissertation presents a detailed analysis of human activity recognition using time series data collected from smartphone accelerometers across six classes of activities. The study employs Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to develop predictive models, aiming to enhance understanding of user behavior and promote healthier lifestyles. The findings highlight the potential applications of this technology in various fields, including healthcare and personal fitness monitoring.

Uploaded by

ace.music2908
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
7 views65 pages

Dissertation Updated Copy - 6607133

This MSc dissertation presents a detailed analysis of human activity recognition using time series data collected from smartphone accelerometers across six classes of activities. The study employs Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to develop predictive models, aiming to enhance understanding of user behavior and promote healthier lifestyles. The findings highlight the potential applications of this technology in various fields, including healthcare and personal fitness monitoring.

Uploaded by

ace.music2908
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 65

UNIVERSITY OF SURREY

Department of Computer Science

MSc Dissertation

A Detailed Analysis Of Time Series Data of Human Activity Recognition over six classes using
CNN & LSTM

A project supervised by Dr Robert Granger

RAJ KANNAN CHENTHIL ARUMUGAM URN:


6607133
I endorse that below is my work and that has not been earlier presented back as a portion of any
assignment. The guidelines on plagiarism stated in the student manual have been understood and ensured
in this report. I have further combined all references implemented in this project at the end.

-RAJ KANNAN CHENTHIL ARUMUGAM


Abstract

Mobile gadgets have become an increasing number of state-of-the-art and the cutting-edge technology of
clever cellular telephones now consists of many numerous and effective sensors. These sensors consist of
GPS sensors, imaginative and prescient sensors (i.e., cameras), audio sensors (i.e., microphones), mild
sensors, temperature sensors, course sensors (i.e., magnetic compasses), and acceleration sensors (i.e.,
accelerometers). The availability of those sensors in mass-advertised conversation gadgets creates
interesting new possibilities for facts mining and facts mining applications. In this paper we describe and
examine a device that makes use of phone-primarily based totally accelerometers to perform hobby
popularity, an assignment which entails figuring out the bodily hobby a person is acting. To enforce our
device we gathered categorized accelerometer facts from 30 customers as they accomplished every day
sports together with walking, jogging, climbing stairs, sitting, and standing, after which aggregated this
time series facts into examples that summarize the person hobby over 10- second intervals. We then used
the ensuing education facts to result in a predictive version for hobby popularity. These findings are
giant due to the fact the hobby popularity version lets in us to advantage beneficial information
approximately the conduct of tens of thousands and thousands of customers passively—simply with the
aid of using having them convey cellular telephones of their pockets. Our findings have an extensive
variety of applications, such as automatic customization of the cellular device’s conduct primarily based
totally upon a person’s hobby (e.g., sending calls without delay to voicemail if a person is jogging) and
producing a every day/weekly hobby profile to decide if a person (possibly an overweight child) is acting
a healthy quantity of exercise.

We are examining 30 individuals that are chosen randomly for this experiment and they analyze the
activity pattern. As an outcome of the research, various purposes, including video monitoring methods,
human-computer interplay, and robotics for individual behavior characterization demand many motion
recognition methods.Human activity comprises movement identification and action model
development. At first, it concentrates on the specific discovery of human actions based upon a
predefined action pattern. Hence, an action memory researcher creates a high-level conceptual
representation and then executes the design by developing a proper pervasive method. By
understanding the problem at hand, we can understand that it is a classification problem. In this project
we’ll cover the life cycle of data analysis starting from importing the necessary packages, data
preprocessing, exploratory analysis of the data, building and validating the machine learning models.
We are also finding the correlation among the attributes in the dataset which helps us to understand the
dependency among the attributes. Secondly, we are developing deep learning models which use
Convolutional Neural Network and Long Short Term Neural Network and compare the efficiency of the
model’s performance. Towards the end, we are finding the confusion matrix and the
classification report for the models developed.
Index

Section 1: Introduction...................................................................................................................................5

1.1 Overview.................................................................................................................................................5

1.2 Project Motivation..................................................................................................................................6

1.3 Project Objectives:..................................................................................................................................6

2. Literature review...................................................................................................................................6

2.1 Need to track human activity..............................................................................................................6

2.2 How can we track the movement of human activity?.........................................................................8

2.3 Challenges faced in tracking human activity?..................................................................................11

2.4 About the data...................................................................................................................................18

2.5 Convolution Neural Network (CNN)................................................................................................19

2.6 Recurring Neural Network (RNN)....................................................................................................20

2.7 Long short-term memory (LSTM)....................................................................................................20

2.8 CNN+LSTM.....................................................................................................................................21

2.9 Risk Assessment:..............................................................................................................................23

3. Methodology and Findings.............................................................................................................24

3.1 Importing libraries and loading data.............................................................................................26

3.2 Data Pre-processing......................................................................................................................28

3.3 Finding the Correlation.................................................................................................................34

3.4 Exploratory Data Analysis............................................................................................................35

3.5 Machine learning models..............................................................................................................38


3.6 Comparison of model evaluation and test results.........................................................................42

3.7 Comparison of Accuracy..............................................................................................................48

4. Conclusion...................................................................................................................................49

References....................................................................................................................................50
Section 1: Introduction

1.1 Overview:

Human action identification or HAR implies a wide range of studies that helps in recognizing the act or
behavior of a character based upon the data
The changes are usually the movements that are conducted inside, such as walking, speaking, holding,
and resting
Human action-perception represents a vital part of human-to-human communication and
interpersonal connections because it presents a knowledge regarding identification of a character,
their nature, and emotional time because of its difficulty to obtain.
The person can understand another human’s quest is one of the central themes of the research since it
includes the fields of computer vision and machine learning. Similarly, an outcome regarding the
analysis, in the form of video monitoring methods, human-computer interplay, and robotics for cultural
manner characterization requires different movement identification methods.

Individual’s movement study comprises human motion identification and action design process. The
primary focus implied here is the specific exposure of social actions based upon a predefined project
design. Hence, an action recognizing a researcher creates a high-level conceptual design at the
beginning and later executes the design over developing a proper pervasive method.
In this project, we are finding the correlation among the attributes in the dataset which helps us to
understand the dependency among the attributes.
We are further refining machine learning patterns by using deep learning methods to classify actions and
analyzing the efficiency of the machine learning model’s execution.
Furthermore, we continue to locate the confusion matrix and the classification report for the models
developed.
1.2 Project Motivation:

The motivation of the project infers to obtain the most suitable fit design toward the supplied train and
test dataset about human activity recognition. Before we construct the machine learning models, we aim
to find a specific correlation between the attributes within the dataset. The machine learning models
developed are CNN , LSTM , Bi directional LSTM and CNN+ LSTM.

1.3 Project Objectives:

This project has 2 major objectives

• Determining the similarity between attributes present within the dataset

• Determining the most suitable fit model for the given train and test data for human activity
recognition.

2. Literature review:

2.1 Need to track the human activity:

With the advent of artificial intelligence into diverse fields especially in mobile computing and desktop
computing devices had become smart, and these smart technologies can be incorporated into user’s daily
life and can be embedded into these devices to track the user’s activity and guide them to be in a good
direction to maintain healthy lifestyles. Our motto behind this project is to mostly encourage, remind and
guide users in a good direction of a healthy lifestyle.
As there is a famous saying ‘health is wealth’ In this generation of the millennium, people had the luxury
to get everything delivered to their door with just a movement of fingertips thereby discouraging users
from involved in any kind of physical activities and on top of that the variety of junk foods that are
readily available to be delivered are making it more calorific intake for the people than been burnt down,
and the outcomes vary from small health glitches like obesity to some serious anomalies like cardiac
arrest etc., It is largely applied in health schemes too that are introduced in the residential surroundings,
hospitals and rehabilitation centres. It also continues to be extensively practiced by observing the actions
regarding the old people staying at rehabilitation stations for prolonged illness control and infection
blocking. This notebook is to build a model that can predict the actions of the human.

Due to the progression of microelectronics and machine systems, sensors are presently developed in
most smartphones that can actively observe our everyday entities besides, observing the essential signs
of patients which have grown on to be one of the highlighted analysis domains in the last few years.
Further, this led to an improved living prospect and a rise in ageing people across several nations, but
due to lack of proper medical care in rural areas has resulted in the need for observing the physical
activities of the patients.

Today’s serious illnesses pose a query to the medical society. An enormous number of studies
concentrates on human movement identification which can assist to determine the relationship amidst
common life including specific research. The learning can be considered essentially a base for planned
medical questionnaires.

Added strategy is the novel energy-efficient method that is practiced toward the identification of
individual movements that target aided living purposes essentially as remote patient motion monitoring
the disabled plus the old.
Remote patient monitoring now signifies by recognizing disabled and ageing patients’ endless strength
while they conduct formulated actions during the day. New population benchmarks prove that the planet
people are ageing quickly.
As an instance, the predictions concerning modifications in people’s bodies through gatherings in Europe
continue pointing that within 2060 the aged (particularly someone above 65 years) would be nearly 30%
of the community. Here illustrates an ominous rise i.e. higher than 70% of the age group and further
presenting unique difficulties over the review station, which intends to obtain effective advantages for
assuring strong existence toward the people.

This study proposes to use the Human Activity Recognition dataset that is openly ready in a new pre-
processed setup besides the aides of the widely used tools available at affordable costs and advanced
machine-learning approaches that quickly monitor the data and can assist as a principle in ascertaining
equitably.

2.2 How can we track the movement of human activity?

Human action recognition intends to match a person's progress of a group caught via sensors.

Presently, accumulating the model like data is not a difficult task. Amidst the growth regarding the
Internet of Things, nearly everyone has a type of device that observes their steps. It can denote a
smartwatch, one pulsometer, or any smartphone.

Consistently, this is managed over heeding a fixed-length sliding windowpane method toward the
characteristic’s removal where a couple of parameters needs to be adjusted: the size of the window and
the shift.
Below are some of the information you could practice:

• Body speed up (acceleration)

• Gravitation acceleration.

• Body crossing velocity.

• Body lean expedition.

• Body slow down (deceleration)

The machine learning pattern is practiced for motion identification and it is formed on the top of these
designs open sensors. However, due to some complexity regarding individual movements including the
current variations within two people, interpreting the data can remain a huge hurdle.

Advantages

• Movement identification is the reason for the growth of various possible purposes in well-
being or sports.

• Examine every movement about a character from the data acquired through various
means.

• Discover what holds as a significant variable that decides the movement in creating a body.

• Determine an imminent design that can identify a person's movement of the signs collected
through the sensors.

• Plan the individualized activity reports improving the well-being of each human.
To implement a smart health care system a unique portion concerning the ageing community disapproves
of age-allied health difficulties in the form of cardiovascular disease, diabetes, Alzheimer’s condition,
osteoarthritis, dementia, or modified tenacious ailments.

These common ailments, mutually by the gradual weakening in cognitive including physical
capacities of ageing characters limit them from surviving at ease. Modern advancements in the ICTs
named as information and communication technologies alongside developments in technologies,
such as assessors, smartphones have endured a swift growth concerning active background. Smart
healthcare seemed to imply a hopeful resolution over the request about an expanding ageing people.
People can strive for smart health assistance to meet the necessities of the expanding community.

Particularly, smart healthcare policies observe and evaluate any significant strength form of aged people
in their everyday actions. The smart health care structures do not just allow the aged people to survive
individually, but then they shall implement healthcare settings that are more sustainable by reducing the
weight toward the entire health policy via the ageing and dependent people at a further stage.
To achieve a smart health care scheme, there exist numerous difficulties in various features regarding
the growing means. Certain tests include the distant examining of the circumstances and the required
IT technology regarding the ecosystem and also the presence of smart processing methods, plus the
presentation of context-aware duties. Hence, to reduce certain difficulties, previously mentioned ones
as well the extra research is required to develop the model.
2.3 Difficulties encountered in tracing human activity?

Human Activity Recognition (HAR) strives towards identifying the movements over individuals or
discussions within groups and perception about learning regarding the user environment and its
enclosing surroundings. Further, this is caused through exploitation regarding the environment along
with the on-body sensors and the allocated computing sources.MC-HF-SVM grants altering fixed-
point by representing a particular description of the number of bits to check the cover model efficiency
and its complexity, leading to advancements in both the aspects of identification and precision

Different methods have been earlier stated within the literature to the identification of human actions
including the different purposes of areas mentioned as healthcare, smart homes, universal computing,
ambient aided subsistence, monitoring and safety. These methods can continue to be classified according
to various diverse measures: by the sensor, which is reliant on covering the signs marked In interest
towards the basic telephony and by including multitasking plus a mixture of sensors that has a
combination of certain portable devices within our everyday life that signifies to expand rapidly and
further transpires to be visualized.
Such methods can seamlessly observe and maintain a record regarding our motions and aid us in
producing results. Such assistive skills can comprise a tremendous significance for distant fitness
control, concerning specifically the ageing as well as the disabled including the ones with
requirements, if they remain independent and smart.

Currently, there exists a real ability toward obtaining the data amidst the active designs. There also
remains a restricted capacity concerning an impulsive result in phases that help in producing insight
about the huge data depository.
Additionally, there remains an instant requirement for new data mining and machine learning methods
to be progressed to this endpoint.
Here, we recommend a unique design concerning social motion identification employing smartphone data,
with the help of automated aided subsistence technologies.
Activity recognition policies strive to recognize these activities conducted down through a person, of
the results obtained from sensors plus its enclosing surroundings

The prevailing smartphones possess movement, acceleration or inertial sensors, employing a particular data
regained from those sensors.
Classification of movements including performances can continue to be identified. Self-regulating
perceptions regarding actions and results are likely via preparing the sensor information by relevant
machine learning and data mining strategies.
The remainder of the report remains classified by the above strategies.

The portions of the unrestricted movement recognition dataset are utilized within this work defined
through sections that explain the suitable backdrop activity executed within the arena.
Moreover, the suggested self-regulating movement recognition method is reviewed and the empirical
validation of the recommended method is defined, hence this report ends after gaining results acquired
through the work including the unique plans for further study.
It is unreliable to entail in traditional home environments for humans to use them at a regular period and
practice them during their activities since the complexity, period and adaptability are to be trained at a
consistent phase to improve the recognition display.

Smartphones possess an edge because of their efficiency and then the availability, besides the capacity of
various sensors on the phone, which can remain utilized towards moment recognition.

Related machine learning and data mining techniques are essentials when they are acquired towards
preparing sensor signs of smartphones toward spontaneous and productive moment identification.
However, they are ought to be different machine learning techniques available and it is not fair, for not
knowing the better procedure that can be helpful for motion detection via smartphones.

If automated movement recognition methods can be formed based upon creative processing of several
sensors traits on smartphones, then the situation will signify a vast addition toward the eHealth domain,
especially towards the remote action monitoring and identifying in the aged care and disability charge
section.

In this report, by the growth of smart devices and smartphones including various sensors, here exists a
leaning towards the HAR and other functions towards receivers, considering inspection, self-overseeing
the body and wellness.
Nevertheless, the foremost challenges regarding the present HAR are the ones that have an efficiency
that is partially low and counting power is required.
Knowing people’s actions including their links by the natural state is a significant element for the
development of devices and digital powered requests. HAR exists as an observation region that
accurately handles the dilemma towards identifying and stimulating. Having in mind the purpose to
dispatch needful learning that can be engaged to provide customized assistance for various purposes.

The HAR has transformed into healing, defense, and service purposes. Therefore, identifying human activities
in normal day to day survival, for instance, walking, running, sets out to be extremely valuable to provide
critique toward the
Caretaker regarding the patient’s condition. Thoughts on sensor-based HAR must remain evolving in the
preceding decade. Some techniques have been intended to identify human’s everyday actions.

One study associated deep learning methods with traditional pattern identification requests aimed at
sensor-based action classification. Thus, the new approach of deep learning method has remained
affiliated but also examined the sensor-based action identification duties of three divisions: sensor
modality, deep model, and applications that ̈ introduced a plan toward sensor-based HAR that can be
updated & combined efficiently towards particular sensors within mobile
Devices. This structure possesses a couple of components: primary, a smartphone aimed at transient data
warehouse registered at different sensors including their software interface and then they address the
information towards the primary server. Following this, a learning strategy is operated to interpret sensor
data to obtain relevant learning for the HAR.

The outcomes showed that the benefit of the design was demonstrated in an actual period where they
combine sensors from various wearable accessories such as smartphones & smartwatches. Lately,
several attempts have been performed via acknowledging human action employing a variety of
suggested methods that helps to manage solely in predefined environments by providing modern
solutions.

Real-time identification applications anticipate users upon classifying adaptable sensors that dispense
via possible contradictory circumstances basically as damaged or failed sensors. In sequence towards
the implementation of Human activity identification employing machine learning techniques

The sensors are employed in HAR and should be inferred to some amount of the original system
support. Different ways are suggested to reduce the measurement components and wearable device
components, counting their place in the human body. The recommended sensor collection system selects
a set of procedures aimed at managing devices positions below various situations. Various algorithms
were analyzed to identify transient models from single activities for displaying multiple actions. Most of
the study is directed on the identification of simple steps to achieve a pure desire identical to walking
and resting alone, whereas the action of many activities across today stands
Considerably more illustrative of an individual’s actual life. The result declares that the exactness
regarding the method can identify multiple actions since temporal forms toward trade include the
inaccuracy. A deep learning method online is recommended. The model was based on various levels
recommended by wearable action identification methods by employing various or homogeneous sensors
to produce sufficient learning.
It is recommended that a sensor-based motion identification method is performed with a multi-sensory
design including an ensemble pruning method. Pair of new pruning techniques are proposed to
estimate the sensor measurement and policy review.

The pruning links to the reduced error and with less error is taken as the final composite classifier via
employing a proper way that remains more reliable than regular group practices. It presents an answer
to manage problems correlated to the growth of size and ageing concerning the community by allowing
monitoring actions of everyday living and modified conditions.

By employing selection techniques, the sensors regarding the smart ecosystem are optimized and greater
efficiency is obtained. The recommended sensor method decreases the prices from a scientific viewpoint,
although the situation preserves the efficiency aimed at motion identification then decreases the sensor
information. Computational intricacy proposed a system that connects the open environments to improve
the liveliness of the identification method in a vibrant atmosphere where the primary input source might
lose, and current data references can remain clear. The preliminary outcomes exhibited that the
completion of the motion perception can remain improved with dynamically determined information
references and the activities are recognized by portable receiver sensors and phone technology
progressing on developing each day, and it remains entirely toward the purpose of explaining
personalities. However, to achieve this, various industries require to be taken out towards the people to
understand their required actions and on how technology is prepared to assist them to complete the tasks.

Once sensors are combined with certain active means, the data reported through certain sensors are to be
examined practicing machine learning processes. HAR has performed a significant part during
developing the growth of people’s state of living

This continues to focus upon healthcare domains plus in a way that HAR can develop ill bodies
by a particular condition towards improvement.
Also, few smartphones devices cannot be altered by the method. So, integration is required to
guarantee that motion identification isn't affected by smartphone energy.

An analysis was carried to manage obesity concerns via HAR through sensors shown in a
smartphone, to observe the user’s dynamic actions, as walking, jogging, sitting, etc. HAR travels into
the method of analysis, fundamentally as base plus meta-level classifiers. Still, HAR commonly
adopted the support level, while placing the classifiers toward the connection approaching a mixture
of action perception duties, reveals us the meta-level classifiers outperforms the base-level classifiers

Another analysis problem is where smartphone users manage to retain their methods inside a secured
manner, (pockets). Therefore, this performance changes the precision. Consequently, they introduced a
scheme associated with parameters correction corresponding to the point of smartphones. That way the
system can improve the execution of the identification of the movement.

HAR denotes a method applied to examine human movement and performance. HAR can show unique or
multiple associates. The sensors are linked in observing users through various centers so as a camera.
Special sensors remain flexible and limit the flow to follow familiar activities. One of the limitations is
the crucial information from the enormous amount of data which is challenging to obtain.
Parallelly, another limitation is to study the problems and situations of efficiency level when the
consequences can be allowed or refused. Still, the action identification is not solely on IT conditions
though it can denote via the researcher to improve the well-being support.

Through the growing quantity of wearable devices and cell transceivers, HAR possesses to draw several
researchers via automatically distinguishing human response including present specific learning upon
their actions, on behalf of active houses setting for wellness & observing support applications, pressure
settings, and shipping of support aids.
The existing research explained that HAR is built upon the sensor information remains and is extremely
stimulating, indistinct by the occurrence concerning a variety of machine learning methods. Here is no
separate answer and overall if it gets towards machine learning practices, that is displayed during the
research methods cannot be a part of the analysis methods since they remain implemented well,
subsequently, it is necessary to periodically examine and analyze each method. Furthermore, it remains
likewise essential in continuing HAR to different actions such as relaxing or bicycling then secure
further characteristics that might assist in explaining human-to-human communication plus relational
connections. Despite the current growth of a few methods such as health informatics, machine learning,
and data mining one shouldn't overlook that the particular personal performance isn’t solely inevitable,
however, an individual can conduct many actions by the corresponding period or also conduct specific
actions that may not relate. An added problem is the difficulties faced in predicting the activity of the
motion, or action.

We believe that the upcoming HAR should be created to foretell and understand these collective actions
and be proficient in managing a chance to gain great efficiency and better healthcare functionality.
The idea of being capable to analyze what action a character is experiencing on a given period is to
enable machines to render support and direction to a character before or while offering a task. The
problem rests in how different the fields of our progress are seen in our day-to-day duties.

The Risk of Falls

The long-term purpose of the information is staying gathered to analyze and alleviate the uncertainty of
bedside acquired by the aged practicing lightweight sensors. The subject of lapses in these venerable has
implied a living one in modern years that the demographics have switched toward an ageing community.
The Centers for Disease Control and Prevention advises that each year, 29 million older people are
dropping with 2.8 million of these cases handled in Accident and Emergency Departments. These
circumstances can include the losses of 28 thousand American citizens periodically. While this is the
worst-case situation, there are various potential consequences of a decline, varying from bone breaks to
traumatic brain damages to depression. Earlier such reviews had remained based around analysis
problems consuming more substantial battery-powered sensors strapped. The receivers need support
while the remaining are expensive and dense.

Therefore, people have been determined to be offensive when they are required to carry them and have
been considered ineffective concerning the issue in subject.

2.4 About the data:

The data was obtained with smartphones fixed at participants’ waists. The purpose of the research
means to examine a dataset of individual actions of around 30 volunteers and seek to examine the same
and produce insights. We further seek to recognize if we could classify the members from their
walking habits and strive to bring further insights. These potentials can be utilized in situations like
activity detection, observing personalities for symptoms of exhaustion, separating one individual from
another, by potential deployment in deeply delicate and safe workplaces etc.

To achieve the dataset for model training, trials ought to be transmitted via society concerning 30
acquaints in an age section between 19 to 48. Individually people executed six activities such as walking,
speaking, holding, and resting with a smartphone upon their waistline. Applying its entrenched
accelerometer and gyrostat, we caught three axial long acceleration and three axial inclined speeds by a
consistent speed of fifty hertz. The preceding methods must be documented to specify the information
manually.
The acquired dataset holds randomly separated into a couple of sets, where 70% of these acquaints are
sustained to produce the training records including 30% of the test information. The data was a raw time
series data.

The obtained data was used in the data transformation process where files of raw accelerometer data are
converted to Attribute-Relation File Format (ARFF files) which was used in modelling.

For our transformation process, we take 10 seconds worth of accelerometer samples (200 records/lines
in the raw file) and transform them into a single example/tuple of 46 values. Most of the features we
generate are simple statistical measures.

Things to note:
An error concerning the number of tuples saved was recently found and corrected in the source code, so
this particular version of the JAR file is not the same one used to create the transformed data from the
raw data that is currently published on our site.

During the transformation process, only the first character of the activity label from the raw data files
are used when creating the arff files. Because some of our activities begin with the same letter (i.e.
Stairs, Standing, Sitting) if these labels are present in the raw files and the JAR file is called, one cannot
distinguish between the activities in the arff files because the activity label will be the same for multiple
activities. WISDM uses a single-character labeling system to represent the activities we recognize, and
simple perl scripts are called when it is necessary to translate between the full activity label and our
single character system.

2.5 Convolution Neural Network:

In the past few decades, Deep Learning has proved to be a very powerful tool because of its ability to handle
large amounts of data. The interest to use hidden layers has surpassed traditional techniques, especially in
pattern recognition. One of the most popular deep neural networks is Convolutional Neural Networks.

CNN’s were first developed and used around the 1980s. The most that a CNN could do at that time was
recognize handwritten digits. It was mostly used in the postal sectors to read zip codes, pin codes, etc. The
important thing to remember about any deep learning model is that it requires a large amount of data to train and
also requires a lot of computing resources. This was a major drawback for CNNs at that period and hence CNNs
were only limited to the postal sectors and it failed to enter the world of machine learning.

The availability of large sets of data, to be more specific ImageNet datasets with millions of labeled images and
an abundance of computing resources enabled researchers to revive CNNs.

Convolutional neural networks are composed of multiple layers of artificial neurons. Artificial neurons, a rough
imitation of their biological counterparts, are mathematical functions that calculate the weighted sum of multiple
inputs and outputs an activation value. When you input an image in a ConvNet, each layer generates several
activation functions that are passed on to the next layer.

The first layer usually extracts basic features such as horizontal or diagonal edges. This output is passed on to
the next layer which detects more complex features such as corners or combinational edges. As we move deeper
into the network it can identify even more complex features such as temporal and spatial pattern changes over
the progression of data.

Based on the activation map of the final convolution layer, the classification layer outputs a set of confidence
scores (values between 0 and 1) that specify how likely the image is to belong to a “class.” For instance, if you
have a ConvNet that detects cats, dogs, and horses, the output of the final layer is the possibility that the input
image contains any of those animals.

Similar to the Convolutional Layer, the Pooling layer is responsible for reducing the spatial size of the
Convolved Feature. This is to decrease the computational power required to process the data by reducing the
dimensions. There are two types of pooling average pooling and max pooling.

There are many methods for time series classification. Most of them consist of two major stages: on the first stage
you either use some algorithm for measuring the difference between time series that you want to classify
(dynamic time warping is a well-known one) or you use whatever tools are at your disposal (simple statistics,
advanced mathematical methods etc.) to represent your time series as feature vectors. In the second stage you use
some algorithm to classify your data. It can be anything from k-nearest neighbors and SVMs to deep neural
network models. But one thing unites these methods: they all require some kind of feature engineering as a
separate stage before classification is performed.

Fortunately, there are models that not only incorporate feature engineering in one framework, but also eliminate
any need to do it manually: they are able to extract features and create informative representations of time series
automatically. These models are recurrent and convolutional neural networks (CNNs).

Using CNNs for time series classification has several important advantages over other methods. They are highly
noise-resistant models, and they are able to extract very informative, deep features, which are independent from
time.
Imagine a time series of length n and width k. The length is the number of time steps, and the width is the number
of variables in a multivariate time series. For example, for electroencephalography it is the number of channels
(nodes on the head of a person), and for a weather time series it can be such variables as temperature, pressure,
humidity etc.

The convolution kernels always have the same width as the time series, while their length can be varied. This
way, the kernel moves in one direction from the beginning of a time series towards its end, performing
convolution. It does not move to the left or to the right as it does when the usual 2-D convolution is applied to
images.

The elements of the kernel get multiplied by the corresponding elements of the time series that they cover at a
given point. Then the results of the multiplication are added together and a nonlinear activation function is applied
to the value. The resulting value becomes an element of a new “filtered” univariate time series, and then the
kernel moves forward along the time series to produce the next value. The number of new “filtered” time series is
the same as the number of convolution kernels. Depending on the length of the kernel, different aspects,
properties, “features” of the initial time series get captured in each of the new filtered series.

The next step is to apply global max-pooling to each of the filtered time series vectors: the largest value is taken
from each vector. A new vector is formed from these values, and this vector of maximums is the final feature
vector that can be used as an input to a regular fully connected layer. This whole process is illustrated in the
picture above

2.6 Recurring Neural Network (RNN)


Recurrent neural networks (RNNs) are deep learning models, typically used to solve problems with
sequential input data such as time series.

RNNs are a type of neural network that retains a memory of what it has already processed and thus can learn
from previous iterations during its training.

A recurrent neural network (RNN) is a class of artificial neural networks where connections between nodes
form a directed graph along a temporal sequence. This allows it to exhibit temporal dynamic behavior. Derived
from feedforward neural networks, RNNs can use their internal state (memory) to process variable length
sequences of inputs …. Recurrent neural networks are theoretically Turing complete and can run arbitrary
programs to process arbitrary sequences of inputs.

A neural network — of which recurrent neural networks are one type, among other types such
as convolutional networks — is composed of three elementary components: the input layer, the hidden layers,
and the output layer. Each layer consists of so-called nodes (aka neurons).

 the temporal lobe of our brain => artificial neural networks => mainly for classification and regression
problems => one of the functions of the temporal lobe is long-term memory

 the occipital lobe => convolutional neural networks => mainly for computer vision problems
(though temporal convolutional networks, TCNs, can be applied to time series)

 the frontal lobe => recurrent neural networks RNN => mainly for time series analysis, sequences, and lists
— for instance, in language processing, which deals with sequences of characters, words, and sentences
ordered by a grammar; or time series, which consist of temporal sequences of observations => one of the
frontal lobe’s functions is short-term memory

A recurrent neural network, by contrast, retains a memory of what it has processed in its recent previous steps.
It makes recurrent connections by going through temporal feedback loops: the output of a preceding step is used
as an input for the current process step.

The hidden layers are located between the input and the output layer. In an RNN, they not only produce an
output, but they also feed it back (“backpropagate” it) as an input for training the hidden layer on the next
observation. They carry out the training by adjusting the synapse weights throughout the neural network. The
network recalibrates the weights for both the current and the previous inputs, multiplies the vector of input
values with the vector of new weights (thus raising or lowering their importance with respect to the training goal
of lowering the prediction error), and passes the vector of results on as an input to the next layer. By adapting
the weights, the hidden layer incrementally derives a kind of function which transforms the input values to
output values that approximate the actual observations in the training dataset. But the function that maps the
inputs to the outputs is not expressed as a closed-form equation — it remains hidden.

If the RNN deals with time series, each period will be represented by a node, holding the period’s observational
value.

An input value, when it is passed from its node on one layer to a node in another layer, travels along an edge (the
connecting line) between the nodes. The edges are comparable to a brain’s synapses.

The receiving node sums all the inputs it receives to a total net input.

It feeds this net input into an activation function (aka transfer function) to compute the output: how much the
node will contribute to the next layer. Among the types of activation functions in frequent use, you will find the
logistic or sigmoid function; the step or heaviside function (comparable to a Boolean or binary yes/no decision);
the hyperbolic tangent function (tanh); and the ReLU function (rectified linear unit, max(0,x)). Nonlinear
activation functions like the logistic or hyperbolic tangent functions help the network to adapt itself to nonlinear
problems when mapping the input to the training output.

The activation function’s output value is multiplied by a weight factor when it is sent out along an edge to a node
in another layer, or over multiple edges to more than one receiving node.

The calibration mechanism a recurrent neural network applies is described by two terms you will encounter in
most RNN descriptions:

 backpropagation through time (BPTT) aka temporal loops;

 gradient descent.

When the network generates prediction values, it also computes the prediction error, the deviation from the
training dataset. The network seeks to minimize the error by adjusting its internal weights during training.
Backpropagation calculates the partial derivatives of the error with respect to the weights. Then the RNN
recalibrates the weights, up or down, based on the partial derivatives.
The term gradient descent refers to the search for a global minimum by evaluating the partial derivatives . The
repeated adjustments of the weights, down the descent towards a minimal error, will move the model towards an
incrementally reduced prediction error.

A simplified sequence of the RNN’s process steps goes as follows:

 It does a forward pass and computes the prediction errors to obtain the loss values on the training dataset and
on the validation set.

 It calculates the gradients at each layer, and backpropagates the errors, back across t timesteps. Then it
updates the weights; and loops around to another forward pass.

The fitting process of the RNN, so that it minimizes the deviation of its output from the actual observations in the
training dataset, is called mapping.

The cost function (aka loss, error, or objective function) measures the prediction error of the network as a
number, a scalar. The goal of the RNN is to minimize the cost function. Prediction accuracy metrics such as the
mean squared error MSE or root mean squared error RMSE, which we know from other time series methods, can
serve as cost functions.

The network computes the difference between the value of an output node and its corresponding actual
observation. It condenses the individual discrepancies — these local, node-specific errors — in the cost function
to obtain the network’s total error or global error, the result of the cost function.

Then the RNN investigates how the total error is spread out over all the weights in the network. It finds their
individual contributions to the total error by calculating partial derivatives, aka gradients (‘slopes’). The loss
function includes a multitude of weights. The partial derivative with respect to a certain weight reveals how that
particular weight affects the total error. The network varies a single weight, records its effect on the total error and
thus obtains its gradient. This is as computationally intensive as it sounds, therefore the long training phases a
neural network typically requires. Gradient descent denotes the search for the global minimum, the set of weights
that will minimize the total error.

The gradients break the total error up so that the RNN can redistribute the chunks back to every weight that
contributed a chunk. This is backpropagation.
2.7 Long short-term memory (LSTM)

LSTMs are an extension of earlier RNNs, capable to retain a long-term memory and use it to learn patterns in
longer sequences of source data. RNN’s could retain a memory, but only about the process steps in their
immediate past. The LSTM, by contrast, introduces loops that can generate long-duration gradients. It can hold on
to long-term patterns it discovers while going through its loops. At each time step, it can tie together three pieces
of information: the current input data, the short-term memory it receives from the preceding cell (the so-
called hidden state), and the long-term memory from more remote cells (the so-called cell state), from which the
RNN cell produces a new hidden state.

The long-duration gradients resolved a problem called vanishing gradient descent, when the model stops
learnings because the gradient’s slope becomes too shallow for the search to further improve the weights. This
happens when many of the values involved in repeated gradient calculations are smaller than 1. The opposite
problem, exploding gradients, arises when many values exceed 1 in the repeated matrix multiplications the RNN
carries out. Exceedingly large gradients would ultimately cause an RNN to be unstable. The vanishing gradient
problem limits an RNN’s memory to short-term dependencies, whereas the LSTM’s formulation keeps the
gradients steep enough so that the search does not get stuck in a dead-end. The vanishing gradient problem can
surface if the model has to deal with long time series extending over hundreds of periods. Each period will have
to be reflected by a network layer; but deep networks with numerous layers involve a long sequence of matrix
multiplications.

A cell in the LSTM is said to be “gated”. Information is selectively added or removed through gates. The cell acts
like a sieve that determines how much incoming information is captured and how much of it is retained. The
model can decide whether it opens an input gate to store information, reject and delete it from long-term memory
(forget gate), or passes the information on to the next layer (output gate). The RNN makes these decisions based
on the importance weights it learns to assign to the information when it seeks to minimize the error and paces
through its temporal loops. The gates carry out matrix multiplications between the information values they receive
as their current inputs, from short-term or long-term memory. Over time, the LSTM learns which information
pieces are effective in reducing the prediction error; it will open and close gates accordingly by assigning higher
or lower weights between 0 and 1 to the information values. Through its loops, it will let the useful values, with
higher weights, pass through the output gate to form a new short-term memory, but it will discard the low-
weighted values.
2.8 Convolution Neural Network (CNN)+Long Short-Term Memory (LSTM)

A convolutional neural network is one that applies a process known as convolution in determining the
relationships between two functions. e.g. given two functions f and g, the convolution integral expresses how the
shape of one function is modified by the other. Such networks are traditionally used for image classification, and
do not account for sequential dependencies in the way that a recurrent neural network is able to do.

However, the main advantage of CNNs that make them suited to forecasting time series is that of dilated
convolutions - or the ability to use filters to compute dilations between each cell. That is to say, the size of the
space between each cell, which in turn allows the neural network to better understand the relationships between
the different observations in the time series.

For this reason, LSTM and CNN layers are often combined when forecasting a time series. This allows for the
LSTM layer to account for sequential dependencies in the time series, while the CNN layer further informs this
process through the use of dilated convolutions.
The model prepared has a CNN and LSTM stacked over each other to overcome the shortcomings of them,
individually. We see a CNN+LSTM model outperforms an only CNN and an only LSTM models.
2.9 Risk Assessment:

Risk assessment is done by using Moss & Atre’s method,

Yellow
Variable Green (Low) (Medium) Red (High)
Technology Skilled thru Marginal Latest
matured practice thru expertise -
expertise skill diminutive
experience
Complexity Effortless, Modest, roughly Task-crucial
nominal system system effect shall entail
effect widespread re-
production

Integration Stance- Restricted Wide-ranging


unaided, no integration integration
assimilation mandatory essential
detected
Ethical Succeeding the Marginal Ethical Plagiarism
issues code of integrity threat meant for exists
from ACM & non-
BCS referencing

Project Data-centric, industry- Less possibility


scope industry- driven, aimed at
ambitious talented, no business and
customer- client
centrical

The risk calculation remains green overall, which helps us to understand that the plan is viable to grow,
the crucial advantage we realize after this remains that the project is ethically strong as it follows the
code of ethics from ACM & BCS.
Whereas the complexity factor is yellow since building and evaluating a machine learning model is a complex
task.

3. Methodology and Findings

The software development methodology is the software development process that is followed to
develop, test, and deliver the software within the defined deadlines while meeting quality standards and
user expectations.

It is the process in which people divide the software development work into smaller independent,
parallel/sequential steps to distribute among a team and enable members of a team to work independently
and collaboratively in a parallel manner.

It also includes the requirements gathering and defining the features out of them and designing and
implementation the features.
In this project, the chosen dataset is collected from 30 volunteers and seek to examine their daily
movement activities. The data is collected using different
wearable sensors placed on the volunteer’s body. The sensors pick up six diverse actions: walking,
speaking, holding, and resting, Walking_downstairs & Walking_ upstairs. Every research gathers
information from diverse people, practices a diverse action set, and exploits various valuation methods.
The sensors can pick and capture the human body motion. The recognition of the activities will be noted
by the sensor. The sensor helps the human to track the activities. Human activity recognition is already
partitioned into training and test datasets where 70% of the dataset is given to the training dataset and
30% is given to the test dataset. The programming language used for developing the program in python
and the interface used for development was Google col-lab. Google co-lab’s major advantage is you can
access the code in co-lab from any system thru categorization into your google account. The prerequisite
is in the direction to install the essential packages in the co-lab to use python in co-lab.
3.1 Introducing libraries and uploading the data:
Google co-lab do not hold all libraries that are said to be predefined inside the system. For it to run with
functions, we require it to carry the essential packages and the libraries. Few packages signify the needed
libraries in the plan and they are named- pandas, NumPy, seaborn, sklearn, tensorflow and matplot. lib.
The purpose of the Numpy library is that it has all the linear algebra designs. There are no predefined
libraries present in google co-lab to work by python we require to install the essential packages. Numpy,
pandas, matplot.lib, seaborn and sklearn are required packages for this project. Numpy is a library that
consists of all linear algebra functions; All the data processing functions are used by panda’s library.
When we use co-lab, it is very simple to access the Google drive, subsequently that we can import the
dataset towards the drive besides export it towards co-lab, while using a preeminent file we will have to
access unique support code meanwhile fulfilling each plan will be successful.

The above code snippet will be useful to mount drive into co-lab while using it, it will generate the
folder pathway which “/content/drive” is followed by the folder position in the google drive.
Dataset Observation:

In dataset view, the extensive features regarding the dataset were implied seeming toward the head,
shape, info.

The preceding illustration will help us to find the source from these train data. While using the head
purpose we will be able to know each header regarding every dataset including its features.

Each display snippet shows the test and train data of the result displaying the number of records and
features. The functions explain the count, non-null and
the datatype of the features. The function also shows the range index of the dataset.

3.2 Data Pre-processing


Data pre-processing must be done within each model while analysing the data because the data needs to
be in an inefficient way. The data pre-processing helps to increase the efficiency and performance of the
analysis. In the scheme we halt toward specific null values inside the dataset, we stop concerning the
categorical variables plus we hold toward the statistical variables and the correlation.

The snippet code shows the IsNull function where we can check the figure of null values existing along with
the feature and also existing within the training dataset. While checking it we find no null values existing
within the training dataset.

We have known that there are no null values therefore we are neither required to remove any absent
values.
The snippet code will be helpful for us to determine if some null values continued to be existing within
the test dataset. We have known that the result is displaying as 0 indented at entirely on the features
showing that specific test datasets do not possess any invalid values.
Categorical variables:-

The same features are shared in the train and the test dataset, so we are only analyzing the training dataset for
categorical variables. While analyzing it we have found that we have activity variables.
Numerical variables:-
This snippet picture displays some characteristics which remain regulated upon statistical variables. The
function aids us in the direction to get all these statistical variables inside the dataset.
3.3 Finding the Correlation:-

The code snippet shows the features and absolute correlation. In this dataset, we are comparing and
analyzing feature 1 and feature 2 which will provide the absolute correlation i.e. above 0.8. The absolute
correlation will help us to find the direction and the magnitude of the relationship between two variables.
Here the variables are feature 1 and feature 2. The greater value is helping us to get the efficient one.
3.4 Exploratory Data Analysis:-

The above graph shows us the exploratory data analysis which is to visualize and analyze the dataset in
the training dataset. The graph shows ups and downs variation in all the activities, which will predict the
values as walking, speaking, holding, and resting, walking (Downstairs), walking (Upstairs). The
Walking shows the higher value which is above the predicted value. It will show how the human activity
works based upon certain things. The graph exactly shows us how we are predicting and analyzing the
values. The day-to-day activity and the action performed by humans working upon will result in the
score which is predefined and analyzed in the simplest form.
The above graph shows us the data analysis which is to test the dataset. Furthermore, the above bar chart
shows us the count where Standing is denoted higher. Comparing the human activity on a daily database
we can find the number of Standing is higher on the daily base count. Therefore, while predicting the
value and analyzing the dataset the outcome produces a higher number of values in all the formats
because humans daily need to perform certain activities. Priority to the other higher values, the walking
activity is increasing but the predicted value is high.
The above graph explains the data points in the training dataset. The graph shows those activities which
come under per day activity. In these data points, we must compare the walking, speaking, holding, and
resting, Walking (Downstairs), Walking (Upstairs).
From the chart, we have found that the Laying denotes 1400 counts per activity. The values with regards
to standing are slightly lower than the laying count per activity i.e. 1300. The Walking denotes a count of
1200 per activity and the Sitting denotes a count of 1250.
The Walking (Upstairs) & (Downstairs) denote 1100 and 900 counts per activity respectively.
Comparing all those values for the training dataset, we have found that per day activity is very helpful
for us in predicting.
The above graph shows the data points in the testing dataset. The Standing and laying data are higher
compared to the rest in the dataset. It shows in most cases that humans are conducting a kind of exercise
as mentioned above. The Walking and Sitting datasets are marginally the same but slightly different in a
few day-to- day activities. Humans are good at walking, the testing dataset displays that walking upstairs
and downstairs seems to have a huge difference in these types of activities.

3.5 Machine learning models


Before, we proceed with developing and evaluating the machine learning models. We need to first split
the data into training and test dataset. In this project, our target variable is Activity which is a categorical
variable. Activity has 6 possible outcomes so as walking, speaking, holding, and resting, Walking
Downstairs and Walking upstairs. The splitting of the dataset happens in where y_train & y_test infers to
activity and X_train & X_test drops the activity attribute from them.
The above snippet helps us to find that we have imported some libraries.
The machine learning prototypes are practiced here to train and test this dataset. First here we are going to
train the dataset by using the machine learning model therefore we will evaluate and execute the test
dataset. The training dataset is nothing but to train and find the null values which are compared to the test
dataset. The training dataset will help us to find an easy way to analyze and predict the datasets.
Model Evaluation: -

The evaluation helps us to predict the value and to analyze the training dataset values. The training
dataset is a function that helps us to check the values and it gives the best precision of the prototype.
Further, the training dataset continues to create, and we will split the dataset for testing the dataset.

The training dataset will help the model to predict the outcome of this project. Additionally, from the
dataset, we must find the score, accuracy, confusion matrix and classification.
While evaluating the values we are finding the dataset values which is giving the probable outcome. The
testing dataset helps to check the values in the training dataset. The testing dataset helps us to find the
way to give the exact outcome of the model. We are finding the accuracy, confusion matrix and
classification in the testing dataset.

3.6 Comparison of model evaluation and test results CNN -

In this, we are going to compare the results obtained from the evaluation and test results. For comparison,
we are going to take the Model accuracy, Confusion matrix and classification report. The results are
below,
In the above snippet images, we can see that on the left is the model evaluation and on the right is the test
results. We can see that the accuracy of the model has dropped slightly compared to the evaluation,
where we can see that model evaluation has 0.860 and test results have 0.850.
As we compare the precision scores, we can see that the scores of model evaluation is high compared to
test results. Following, the comparison of recall value test results is higher than the evaluation results.

LSTM: -

The comparison of results from model evaluation and test is done. We are comparing the model
accuracy, confusion matrix and classification report. The results obtained are,
The above images display the results obtained from model evaluation and test results from the decision
tree classifier. We can see that accuracy model evaluation is higher than test results. We can also see that
precision, recall and f-1 score are also higher than test results. This happens with the decision tree when
trained and are evaluated with the same data is the reason for the results of the model evaluation showing
perfect 1.

CNN+LSTM:-

In this, we are going to compare the results obtained from model evaluation and test results of the
CNN+LSTM classifier. We are comparing the accuracy, confusion matrix and classification report.
The results above help us to understand that the model accuracy of model evaluation is higher than model
test results. We can also see that precision, recall and f1 score of model evaluation is higher than model
test results. This happens because of the nature of the K-neighbors algorithm since the model is already
trained using training data and the predictions are almost the same.
3.7 Comparison of Accuracy

Accuracy plays a major role while testing the dataset. It is our main factor to check the accuracy.
Accuracy gives the absolute value while testing the dataset. We have used four classifiers to check and
differentiate the accuracy. They are CNN, LSTM, CNN+LSTM. While testing the data CNN+LSTM has
the highest accuracy compared to another classifier. The CNN+LSTM has 0.984 accuracies in the test
dataset. LSTM is the second most accurate compared to the other two classifiers, here the accuracy is
0.9003. CNN has 0.8564.
Comparison is always needed for machine learning models to get the best and positive test datasets. We
can see that all the model values are higher than 80%. There is much difference among accuracy results
because some of the classifiers give the exact or nearest value and some of the classifiers gives the
relative value, however, all the values are accepted therefore classifier is made to predict the values from
the test data.

4. Conclusion

Human movement identification produces an extensive array of methods because it bears upon
wellbeing. The data were tested from multiple test cases under a set of conditions. At first, merging the
four most suitable classifiers followed by practicing the tally of possibilities designs directed out to mean
the most competent datasets aimed at evaluating human action. The test dataset results have shown us the
efficient and perfect way of outcome for this project. The fundamental notions of the activity study
(essentially as sensors, activity practices, etc.).
We are using the machine learning model so that the experiments and the training dataset are fine-
tuned, it also helps us to look for more profound patterns. The recognition based on sensors leads to
many possible kinds of research and gives the best predictable source of human activity recognition.
We have found that the dataset and test patterns gave us a better understanding and in conclusion, the best
methods of the machine learning model have helped us to find the accurate results of Human activity
recognition.
References

[1] E. Kim, S. Helal and D. Cook, "Human Activity Recognition and Pattern Discovery," in IEEE
Pervasive Computing, vol. 9, no. 1, pp. 48-53, Jan.-March 2010, doi: 10.1109/MPRV.2010.7.

[2] O. D. Lara and M. A. Labrador, "A Survey on Human Activity Recognition using Wearable
Sensors," in IEEE Communications Surveys & Tutorials, vol. 15, no. 3, pp. 1192-1209, Third Quarter
2013, DOI: 10.1109/SURV.2012.110112.00192.

[3] N. Oliver, E. Horvitz and A. Garg, "Layered representations for human activity recognition,"
Proceedings. Fourth IEEE International Conference on Multimodal Interfaces, 2002, pp. 3-8, DOI:
10.1109/ICMI.2002.1166960.

[4] Y. Wang, K. Huang and T. Tan, "Human Activity Recognition Based on R Transform," 2007
IEEE Conference on Computer Vision and Pattern Recognition, 2007, pp. 1-8, DOI:
10.1109/CVPR.2007.383505.

[5] J. Ben-Arie, Zhiqian Wang, P. Pandit and S. Rajaram, "Human activity recognition using
multidimensional indexing," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
24, no. 8, pp. 1091-1104, Aug. 2002, DOI: 10.1109/TPAMI.2002.1023805.

[6] O. C. Ann and L. B. Theng, "Human activity recognition: A review," 2014 IEEE International
Conference on Control System, Computing and Engineering (ICCSCE 2014), 2014, pp. 389-393, DOI:
10.1109/ICCSCE.2014.7072750.

[7] Michalis Raptis, Leonid Sigal; Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2013, pp. 2650-2657

[8] M. A. R. Ahad, J. K. Tan, H. S. Kim and S. Ishikawa, "Human activity recognition: Various
paradigms," 2008 International Conference on Control, Automation and Systems, 2008, pp. 1896-1901,
DOI: 10.1109/ICCAS.2008.4694407.
[9] T. Lan, L. Sigal and G. Mori, "Social roles in hierarchical models for human activity
recognition," 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp.
1354-1361, DOI: 10.1109/CVPR.2012.6247821.

[10] A. Madabhushi and J. K. Aggarwal, "A Bayesian approach to human activity recognition,"
Proceedings Second IEEE Workshop on Visual Surveillance (VS'99) (Cat. No.98-89223), 1999, pp.
25-32, doe: 10.1109/VS.1999.780265.

[11] S. W. Pienaar and R. Malekian, "Human Activity Recognition using Visual Object Detection,"
2019 IEEE 2nd Wireless Africa Conference (WAC), 2019, pp. 1- 5, DOI:
10.1109/AFRICA.2019.8843417.

[12] Aggarwal, J. K., and Ryoo, M. S. (2011). Human activity analysis: a review.
ACM Comput. Surv. 43, 1–43. doi:10.1145/1922649.1922653

[13] Xia, L. (2014). Human activity recognition from 3D data: a review. Pattern Recognition. Lett.
48, 70–80. doi:10.1016/j.patrec.2014.04.011

[14] Stack Abuse. 2021. K-Nearest Neighbours Algorithm in Python and Scikit- Learn. [online]
Available at: <https://stackabuse.com/k-nearest-neighbors- algorithm-in-python-and-scikit-learn/>

[15] Vertica.com. 2021. VerticaPy. [online] Available at:


<https://www.vertica.com/python/documentation_last/learn/KNeighborsClassifi er/>

[16] Uc-r.github.io. 2021. Naïve Bayes Classifier · UC Business Analytics R Programming Guide.
[online] Available at: <https://uc-r.github.io/naive_bayes>

[17] VanderPlas, J., 2021. In Depth: Naive Bayes Classification | Python Data Science
Handbook. [online] Jakevdp.github.io. Available at:
<https://jakevdp.github.io/PythonDataScienceHandbook/05.05-naive- bayes.html>

[18] Ieeexplore.ieee.org. 2021. Layered representations for human activity recognition.


[online] Available at:
<https://ieeexplore.ieee.org/abstract/document/1166960>

You might also like