Human activity detection

Summary

This video discusses the importance and implementation of a human activity detection model in the context of Big Data. Highlighting its significance in medical care and elderly support, the script details the dataset from the UCI repository, which includes eight attributes and eleven activity classes. After pre-processing, feature extraction, and model training using a decision tree, the model achieved over 73% accuracy in classifying activities based on attributes like Z coordinates and tag identifiers. The video concludes with suggestions for improving model performance through time series analysis.

Takeaways

• 🧑‍🔬 Activity detection is increasingly important in medical care, supporting elderly independent living and emergency assistance.
• 📈 Companies like Fitbit and Apple rely on activity detection for health monitoring and safety features, respectively.
• 📚 The data set used for the project was sourced from the UCI Machine Learning Repository and includes eight attributes and eleven classes.
• 👥 The data set was created by recording five people performing various activities over five sessions each.
• 📊 The data set attributes include sequence name, tag identifier, timestamp, date, and x, y, z coordinates, along with the activity classification.
• ⚙️ Pre-processing involved checking for missing values and removing highly correlated activities, resulting in three main classes: lying, walking, and sitting.
• 📈 Feature extraction and analysis identified highly correlated attributes like timestamp, tag identifier, and x, y, z coordinates.
• 📉 Box plots were used to visualize variations in attributes across different activities, aiding in attribute selection for the model.
• 🌐 A 70/30 train-test split was found to be optimal for model training, with 5-fold cross-validation repeated three times for parameter tuning.
• 🔑 The decision tree model was chosen for classification, using attributes like x, y, z coordinates, timestamp, and tag identifier.
• 📊 The model achieved over 73% accuracy in classifying activities, with the Z coordinate being particularly effective for distinguishing lying and walking.
• 🤖 Error analysis showed some confusion between lying and sitting, and walking and sitting, suggesting room for improvement in the model.

Q & A

Why is activity detection important in the field of medical care?

Activity detection is important in medical care because it can support the elderly for independent living and can be a life-saving feature, such as in fall detection systems that automatically alert emergency services if a person falls and doesn't get up for a while.

Which companies are mentioned in the script that rely on activity detection?

Fitbit and Apple are mentioned as companies that rely on activity detection. Fitbit for tracking physical activities and Apple for features like fall detection in their Apple Watch products.

What is the source of the data set used for the human activity detection model?

The data set used for the human activity detection model was obtained from the UCI Machine Learning Repository.

How many people were involved in the creation of the data set and what were they asked to do?

Five people were involved in the creation of the data set, and they were asked to perform a sequence of activities over five times.

What are the eight attributes contained in the data set?

The eight attributes in the data set are the sequence name, tag identifier, timestamp, date, and the x, y, z coordinates, followed by the activity classification.

What was the reason for removing certain activities during the pre-processing stage?

Certain activities were removed during pre-processing because they were highly correlated, such as active researches lying and lying down, which had extremely similar variations in terms of their features.

How many class labels were retained after the pre-processing of the data set?

After pre-processing, three class labels were retained: lying, walking, and sitting.

What was the best train-test split ratio found for the model?

The best train-test split ratio found for the model was 70/30.

Which model was used to classify the data based on the given attributes?

A decision tree model was used to classify the data based on the given attributes.

What attributes were selected for training the decision tree model?

The attributes selected for training the decision tree model were the x, y, z coordinates, along with the timestamp and the tag identifier.

What was the overall accuracy achieved by the decision tree model in classifying the activities?

The decision tree model achieved an accuracy of over 73% in classifying the activities.

What was the main issue identified in the error analysis of the decision tree model?

The main issue identified in the error analysis was that the model sometimes confused activities like lying with sitting about 16% of the time, and walking with sitting about 12% of the time.

How can the performance of the decision tree be improved further?

The performance of the decision tree can be improved by conducting a time series analysis of each attribute, which is likely to increase accuracy.

Outlines

00:00

📊 Human Activity Detection: Importance and Data Overview

The video introduces a human activity detection model developed as part of a Big Data course. It emphasizes the significance of activity detection in healthcare and elderly care, citing examples like Fitbit and Apple Watch's fall detection feature. The dataset used is from the UCI Machine Learning Repository, capturing data from five individuals performing various activities. The data includes attributes such as sequence name, tag identifier, timestamps, dates, and XYZ coordinates, along with activity labels. The video explains the pre-processing steps, which involved removing highly correlated activities to focus on three primary classes: lying, walking, and sitting. This decision was made to retain over 80% of the dataset while simplifying the model's complexity.

05:04

📈 Data Analysis and Model Training Strategy

This paragraph delves into the analysis of the dataset, highlighting the use of box plots to understand attribute variations across different activities. It discusses the selection of attributes for model training based on their correlation and variation. The training and testing strategy involves a 70/30 split and 5-fold cross-validation repeated three times to ensure robust parameter tuning. The chosen model is a decision tree, which uses attributes like XYZ coordinates, timestamp, and tag identifier to classify activities. The model's accuracy in classifying activities such as lying, walking, and sitting is reported, with a focus on the decision tree's ability to differentiate these activities based on the Z coordinate primarily.

Keywords

💡Activity Detection

Activity detection refers to the process of identifying and classifying human activities based on sensor data. In the context of the video, it is crucial for applications such as supporting elderly independent living and health monitoring, as well as for fall detection in wearable devices like the Apple Watch, which can alert emergency services if a person falls and does not get up, potentially saving lives.

💡Data Analysis

Data analysis is the examination of data sets to draw conclusions about the information they contain. The video emphasizes its growing importance in medical care, where activity detection through data analysis can provide insights into a person's health and well-being, contributing to better patient care and monitoring.

💡Big Data

Big Data refers to data sets that are so large and complex that traditional data processing software is inadequate to deal with them. The video discusses implementing a human activity detection model as part of an introduction to Big Data course, indicating the use of large datasets in activity detection and the need for advanced analytical methods.

💡UCI Machine Learning Repository

The UCI Machine Learning Repository is a collection of databases, domain theories, and data generators used for machine learning research. It is mentioned in the video as the source of the localization data for person activity, which was used to create the dataset for the activity detection model.

💡Pre-processing

Pre-processing in data analysis involves cleaning and transforming raw data into an understandable format. The video describes checking for missing values and removing highly correlated activities during pre-processing, which is essential for preparing the data for accurate model training.

💡Feature Extraction

Feature extraction is the process of selecting the most relevant data attributes or features from the dataset. The video explains how summary and visualization of the dataset led to the identification of key attributes such as the x, y, and z coordinates, timestamp, and tag identifier, which were used for training the model.

💡Correlation

Correlation measures the extent to which two variables are linearly related. In the video, the correlation between attributes is plotted to identify which features are highly correlated, such as the timestamp, tag identifier, and x, y, z coordinates, guiding the selection of attributes for the model.

💡Decision Tree

A decision tree is a flowchart-like structure in which each internal node represents a 'yes' or 'no' question, and each branch represents the outcome of that question. The video describes using a decision tree model to classify activities based on selected attributes, such as the x, y, and z coordinates, which helps in achieving a certain level of accuracy in activity classification.

💡Accuracy

Accuracy in the context of machine learning refers to the percentage of correct predictions made by the model. The video mentions achieving over 73% accuracy in classifying activities, which is a key performance metric indicating how well the model can predict the correct activity.

💡Confusion Matrix

A confusion matrix is a table used to describe the performance of a classification model. The video uses a confusion matrix to analyze the error in the decision tree model, showing which activities were often confused with each other and the accuracy of correct predictions.

💡Time Series Analysis

Time series analysis is a statistical technique that deals with time series data or trend analysis. The video suggests that further time series analysis of the attributes could improve the performance of the decision tree model, indicating the potential for enhancing accuracy through deeper temporal data analysis.

Highlights

The importance of activity detection in medical care and elderly support.

Companies like Fitbit and Apple utilize activity detection for health monitoring and safety features.

Introduction of a human activity detection model implemented as part of a Big Data course.

Data set obtained from UCI machine learning repository with over eight attributes and eleven classes.

Description of the data set attributes including sequence name, tag identifier, timestamp, and coordinates.

The significance of the activity label in identifying the performed activity.

Pre-processing steps including checking for missing values and removing highly correlated activities.

Final class labels after pre-processing: lying, walking, and sitting.

Feature extraction and data visualization techniques used in the study.

High correlation found between timestamp, tag identifier, and XYZ coordinates.

Selection of attributes for model training based on their variation with activities.

70/30 train-test split and 5-fold cross-validation strategy for model fine-tuning.

Use of a decision tree model for activity classification based on selected attributes.

Achievement of over 73% accuracy in classifying activities using the decision tree model.

Error analysis revealing confusion between similar activities like lying and sitting.

Potential improvement of model performance through time series analysis of attributes.

Summary of the activity detection model's effectiveness and areas for future enhancement.

Transcripts

hi everyone in this video we talk about

the human accurate detection model that

we implemented as a as a part of our

introduction to Big Data course but

before we get into the implementations

let's first have a look at why activity

detection is important as we can see

from the current trend data analysis is

becoming increasingly important in the

field of medical care activity detection

can be important for the purposes of

supporting the elderly for independent

living companies like Fitbit rely almost

entirely on activity detection while

companies like Apple have introduced

features like fall detection in the

Apple watch products which could be a

life saving feature for example if a

person Falls and doesn't get up for a

while

he's probably injured so the watch

automatically allows the emergency

services for assistance for these

reasons these chosen the localization

data for person activity as our data set

which we obtained from the UCI machine

learning repository for the creation of

this data set five people were asked to

perform a sequence of activities over

five times and the data set contains

over eight attributes and eleven classes

which we'll get into in just a minute

alright let's have a look at the

attributes in the data set the data set

has over eight attributes namely the

sequence name the tag identifier time

stamp date x y&z coordinates followed by

the activity it is classified into as we

can see from the notebook in the right

the sequence name is composed mainly of

normal values from a 0 1 to e zero 5

over here a denotes the name of the

person being recorded while the number

denotes the recording session of that

person the tag identifier helps us

identify which Tag sensor information is

actually being reported in the room the

time stamp tells us the time at which

the recording

while the date gives us the date of the

recording the XYZ coordinates as we know

are simply the XYZ coordinates of that

tag finally the activity label gives us

which activity was actually being

performed let's have a look at an

example instance here is 0 1 is the

sequence name a is the name of the

person performing the activity while 0 1

is the first recording instance of that

activity set the second column the tag

identifier corresponds to the chest type

so we know that this row reports theta

mainly for the chest bag here's the

timestamp and the x y&z coordinates for

the data as we can see this particular

row has been classified into the walking

activity next we get into pre-processing

the data as a part of our pre-processing

we check for missing values and removed

highly correlated activities for example

active researches lying and lying down

had an extremely similar variation in

terms of their features so we removed

these activities and ended up with final

3 class labels namely lying walking and

sitting since these are the most

important activities generally performed

by humans we still managed to retain

over 80% of our dataset we can see the

cleaning being performed in the Jupiter

to the right

next we get into the feature extraction

from our dataset as we can see from the

right we found out the summary and

visualize the data set as follows we

found that lying sitting and walking

comprised a major part of the data set

to see the correlation between all the

attributes we plotted the we plotted

their correlations as can be seen in in

the notebook on the right the

observations from the plot can be seen

below we found out looking at the upper

right triangle that the attributes like

the timestamp tag identifier x y&z

coordinates were highly correlated and

this is what drove our attribute

selection to train our model let's have

a look at the variation between each

attributes activity wise as we can see

from box plots for the activity versus

bag identifiers the timestamp the X the

Y and the Z coordinates we are able to

make out how each are to be varies with

activities thus this gives us a good

sense of which attribute to select for

the purpose of training our model next

we dive into the Train and test strategy

we use for our body after dividing date

the data into several ratios we found

the best plate to be the 70/30 train

test split for our model for the

purposes of fine-tuning our parameters

we used a 5 fold cross validation

repeated three times this helped us to

work with the dependent and grouped data

as in the data set below we can see in

the notebook on the right how the data

was split into train and test groups

let's have a look at the model we use to

classify our data we use the decision

tree model for the purposes of pinning

our activities based on the given

attributes in two sitting falling in

line as seen from the attribute

correlations plotted above we selected

attributes such as the x y&z coordinates

along with the time stamp and the tag

identifier for the purposes of training

our decision tree as we can see from the

decision tree in the notebook to the

right based simply on the Z coordinates

we are able to classify activities such

as lying and walking with accuracy of 61

and 53% respectively all right let's try

and evaluate our model now we were able

to achieve an accuracy of over 73% in

classifying the attributes into

activities let's analyze the performance

of our decision tree based on attributes

like the Z coordinate we were able to

classify the data to large extent next

it used the tagger identify attribute to

be able to differentiate between

activities like sitting and walking

thirdly the y coordinate was used to

differentiate between activities like

lying and walking let's get into the

error analysis of our decision tree as

we can see from the confusion matrix we

were able to predict correctly

activities such as lying sitting and

walking with an accuracy of power 82 71

and 80% however our tree still got

confused between activities such as

lying and sitting about 16% of the time

and walking and sitting about 12% of the

time finally let's summarize what we

just saw we saw how each activity varies

with the attribute mainly through the

boxplots

of the activities

absolutes this also validated our

initial ideas that activities such as

lying would have a very little change in

all three coordinates whereas activities

like walking would have no change in

that chest Z coordinate especially while

the person is walking on a straight line

to improve the performance of our tree

we can further do a time series analysis

of each of these attributes which is

definitely bound to increase the

accuracy thank you for your time and

patience have a good day