Project Instructions
Goal:
The goal of the project is to build a desktop app that can distinguish between ‘walking’ and
‘jumping’ with reasonable accuracy, using the data collected from the accelerometers of a
smartphone.
Description:
The project involves building a small and simple desktop application that accepts accelerometer
data (x, y, and z axes) in CSV format, and writes the outputs into a separate CSV file. The output
CSV file contains the labels (‘walking’ or ‘jumping’) for the corresponding input data. For
classification purposes, the system will use a simple classifier, i.e., logistic regression.
In order to accomplish the goal of the final project and complete the report, the following 7 steps
are required:
1. Data collection
2. Data storing
3. Visualization
4. Pre-processing
5. Feature extraction & Normalization
6. Training the model
7. Creating a simple desktop application with a simple UI that shows the output
Step 1. Data collection
In this step, you need to collect data using your smart phone while ‘walking’ and ‘jumping’. There
are a number of different apps you can use to collect accelerometer data from your smartphone.
As an example, you may use an app called Phyphox, which works on both iOS and Android, and
allows you to output the recorded signals as a CSV file. Other apps would also be acceptable.
Data collection protocol: Recall that when collecting data, the diversity of the dataset will allow
your system to work better when deployed. (a) Therefore, to maximize diversity, each team
member must participate in the data collection process to create a total of 3 subsets (1 per
member). (b) To further maximize diversity in your dataset, the phone should be placed in different
positions. For example, you can place the phone in your front pocket, back pocket, pocket of a
jacket, carry it in your hand, etc. (c) The duration of data collection by each member must exceed
5 minutes. Please note that it is important that you collect a roughly balanced dataset. In other
words, the amount of time dedicated to each user, to each action (‘walking’ vs. ‘jumping’), to each
phone position, and others, should be roughly the same.
2
Step 2. Data storing
After transferring your dataset (all the subsets) to a computer and labeling them, store the
dataset in an HDF5 file. This HDF5 file must be organized as follows:
It is always a good idea to keep the data as originally collected, which is why we have the
structure that we see on the right side of this image. But in order to create a simple AI system,
you need to create separate training and test splits. To do so, divide each signal into 5-second
windows, shuffle the segmented data, and use **% for training and 10% for testing. This new
dataset must also be stored in the HDF5 file as shown on the left side.
Step 3. Visualization
Data visualize is a critical step in the field of data science and will allow you to find issues in the
data early on, and also become familiar with the data that you will be working with. So, in this
step, you will need to visualize a few samples from your dataset (all three axes) and from both
classes (‘walking’ and ‘jumping’). A simple acceleration vs. time would be a good start. But also
think about additional creative ways of showing the data with the goal of representing your
dataset. Provide some visualization for the meta-data for your dataset and sensors too. Don’t
forget to use good visualization principles.
Step 4. Pre-processing
Remember, garbage in, garbage out! Almost any dataset, no matter how careful you were during
collection, will inevitably contain some noise. First, the data will likely contain noise, which may
be reduced by a moving average filter. Second, after feature extraction (next step), try to detect
and remove the outliers in your collected data. Please note that if by removing outliers, the data
becomes too imbalanced, remedy this. Finally, normalize the data so that it becomes suitable for
logistic regression.
Step 5. Feature extraction & Normalization
From each time window (the 5-second segments that you created and stored in the HDF5 file),
extract a minimum of 10 different features. These features could be maximum, minimum, range,
mean, median, variance, skewness, etc. Additional features may be explored as well. After feature
extraction has been performed, you will be required to apply a normalization technique for
preventing features with larger scales from disproportionately influencing the results. Common
normalization techniques are min-max scaling, z-score standardization, etc.
Step 6. Creating a classifier
Using the features from the preprocessed training set, train a logistic regression model to classify
the data into ‘walking’ and ‘jumping’ classes. Once training is complete, apply it on the test set
and record the accuracy. You should also monitor and record the training curves during the
training process. Note that during the training phase, your test set must not leak into the training
set (no overlap between the segments used for training and testing).
3
Step 7. Deploying the trained classifier in a desktop app
The last step is to deploy your final model in a desktop app. For building a simple graphical user
interface in Python, you can use Tkinter or PyQt5 libraries. As mentioned, this app must accept
an input file in CSV format and generate a CSV file as the output, which includes the labels
(walking or jumping) for each window in the input file. Run a demo for your built app in which you
input a CSV file and the app generates a plot which represents the outputs. Once deployed, how
did you test the system to ensure it works as intended?
Step 8. Demo video
Record your screen while running a demo with the created app. The video should feature all team
members and show short snippets of your data collection process, as well as the app in action.
The video should also explain your project in a few sentences. It should be between 1 to 3
minutes.
Step 9. Report
Write a report for the project. The project should contain:
- A title page containing the following:
Course: ELEC292
Project Report
Group Number: _
Names, Student Numbers, and Email Addresses:
Date:
- After the title page, the rest of the document must be in 12 point Times New Roman font,
single spaced, 1 inch margins, and with page numbers in the bottom center of each page.
- Every student must submit a separate copy that is identical to their teammates. This is
done as a signoff, indicating that each member has participated and agrees with the
content. It will also make grading and tracking easier.
- As a rule of thumb, the report should be between 15 to 20 pages including references and
figures.
- Note that where you refer to online sources (articles, websites, etc.) the references must
be mentioned in the reference section of the document (in the end of the document), and
the references should be referred to in the text. Here is a brief description of how proper
citation and references should be used: https://labwrite.ncsu.edu/res/res-citsandrefs.html
- In this report, you must use the IEEE format for references.
- Note that in your report, you must not “copy-paste” text from other resources, even though
you are citing them. Text should be read, understood, and paraphrased, with proper
citation of the original reference.
- Proper editing (grammar, typos, etc.) is expected for the reports.
- The report should clearly describe each step and provide the requested material.
- The report must have the following sections:
4
o 1. Data Collection: How did you collect the data, label it, transfer it to a PC, and
what challenges did you deal with from during the data collection step. How did
you overcome them? Mention all the hardware and software used.
o 2. Data Storing: Provide a full description of the way you stored the collected data.
o 3. Visualization: Provide all the plots that you created for visualization purposes,
and provide appropriate descriptions for each of them. What did you learn?
Knowing what you learn from the plots, if you were to re-do your data collection,
how would you do things differently?
o 4. Preprocessing: Clearly describe the measures you took for preprocessing, and
how it impacted the data (you may use a few plots here too). Why did you choose
the parameters that you did (e.g., size of moving average)?
o 5. Feature Extraction & Normalization: What features did you extract and why?
References may be useful here. Explain the process of feature extraction and
normalization, then justify your choices.
o 6. Training the classifier: Provide a description of the way you trained the logistic
regression model. This section must include the learning curves and your accuracy
on the training and test sets. What parameters did you use here? Justify your
answers.
o 7. Model deployment: This section should include the details of how you deployed
the trained model into a desktop app. Provide screenshots of the GUI you created
along with its description, and justify your design choices.
- At the end of the report, a Participation Report must be added. Please note that the
project should be done together and collaboratively. It is not acceptable for one person to
do the technical work and another to simply write the report. Having said this, a reasonable
division of work is allowed for type up or other simple tasks. At the end of the report,
provide a table that clearly shows which members have been present for and contributed
to each question. Please note that should someone not pull roughly 1/3 of the weight of
the project, they may lose points.
Submission:
The following items will need to be submitted in OnQ:
- 1. Your project report in PDF format
- 2. Your saved HDF5 file in the mentioned format
- 3. The video as described earlier
- 4. Your clean and executable Python code, which contains the code for ranging from (1)
visualization, (2) pre-processing, (3) feature extraction, (4) training and running the
model
Bonus:
Part 1: This part of the final project is not mandatory and
serves as a bonus deliverable, which can gain up to 10
bonus points(out of 100) on your project!
The app that you created, works offline. In other words, the
app is not able to classify activities from your smart phone
in real-time. For the bonus component of the project, our goal is to build a desktop app which can
real-time
5
read the accelerometer data from your smart phone in a real-time and classify it immediately. As
shown in the image above, your smart phone would need to send the accelerometer data to the
app in real-time, and the app would show the class of action (e.g., ‘walking’) in real-time.
Hint: For reading the accelerometer data online, you may use the ‘Enable remote access’ option
of the Phyphox app. By doing so, you will have access to the accelerometer data in a web page.
Then, you may use Beautiful Soup and Selenium libraries to read the data. Alternative ways
include using Bluetooth to send the data to the PC in real-time.
Part 2: In this step, you will have to implement the SVM and Random Forest classifier from scratch
without the use of any existing libraries for the model, such as, scikit-learn. You are free to use
libraries for basic operations, such as, NumPy. Record the accuracy of your implemented models
that you built from scratch on test set. Then, compare the performance of your own implemented
models (without using libraries) with the models implemented using pre-existing libraries. This
direct comparison will highlight the efficacy of your custom-built models versus standardized
library models. Finally, provide overall insights on the comparison of models and explain the
outcomes.
Deliverables for the bonus component:
1. The report should be extended by 3-5 pages. These additional materials should include:
o All the details of how the data was transferred to the PC in real-time
o A description of any changes made to the desktop app and its GUI
o A description of any changes made to the trained classifier
o A general description of how you implemented SVM and Random Forest
from scratch
o A table that shows the accuracy of models on test sets that you implemented and
the models exist in standard libraries
2. The video should clearly show that a person is carrying a phone and the desktop app is
classifying their actions in real-time
3. Your clean and executable Python code
General note: If you attempted anything but could not get it to work, whether for the main part of
the project or the bonus component, you should mention what you did, what is your hypothesis
for it not working, and how things should likely change to make it work, to receive some partial
marks.
Grading:
A 5-point scale will be used for grading different aspects of the project. This 5-point scale will be
as follows:
Quality Grade Definition
Excellent 4/4 Explanations are clear and easy to
understand, complete
Good 3/4 Explanations are lacking a bit of
clarity or completeness, but is
generally in good shape
Average 2/4 Several aspects are missing or
incorrect. There is quite a bit of
room for improvement
6
Poor 1/4 Most aspects are missing or
incorrect
Not done 0/4 The question is not answered at all
The following grading scheme will be used:
The final grade for the project will be calculated out of 100. Up to 10 points for the bonus component will
then be added to this grade (if available). The final score will be multiplied by 0.3 to obtain your project
grade out of 30.
Note: The use of generative AI such as ChatGPT is prohibited in Final Project submission and is
considered a violation of the academic integrity principles of Queen's University. Please note that, we
will check the assignments using the latest AI-content detectors on a random basis
Task Grade Weight
1. Data collection / 4 Completeness/thoroughness, balance, diversity, good
data collection principles
3
2. Data storing / 4 Proper data storage in the specified format, reasonable
train-test splits, no data leakage
2
3. Visualization / 4 Several samples visualized, each class represented,
meta-data visualized, additional creative plots, good
visualization principles
2
4. Pre-processing / 4 Removal/reduction of outliers, removal/reduction of
noise, discussion or remedy of imbalance, normalization,
further visualization of data after pre-processing
2
5. Feature extraction / 4 Identification and extraction of a minimum of 10 different
features, proper
2
6. Training the mode / 4 Proper construction
reasonable results
and training of the model, 2
7. Desktop app / 4 Nice/clean UI, functionality, testing of the system 2
8. Demo video / 4 Proper description and demo of the work, participation
from everyone
3
9. Report / 4 Proper structure, detailed description, high quality
images, writing and editing quality, references and
citations, cover page, division of work statement,
providing everything described under Step 9 on pages
請加QQ：99515681 郵箱：99515681@qq.com WX：codinghelp
