Graduation Project - Accident Fatality Predictor

Team 👨🏻‍💻👨🏻‍💻👨🏻‍💻👩🏻‍💻👩🏻‍💻

• Jaehong Kwon
• XiongFei (Frank) Shi
• Feng Wang
• Olive Sun
• Neha Nayeem

Topic and Description

Topic: Prediction of Accident Fatality using Machine Learning

Using historical KSI (Killed or Seriously Injured) data from Toronto Police Open Data, we aim to predict the fatality of an accident given a certain input (time/weather condition/etc.).

This information could be used to possibly prevent accidents by encouraging drivers to drive more carefully in a particular area.

Datasets 🗃

• Toronto Police Open Data

Project Infrastructure

Our app is deployed using Heroku and is available at: Accident Fatality Predictor

Tools used

• Python and Jupyter Notebook
• Libraries: pandas, numpy, sklearn, pickle, keras, tensorflow
• Flask
• Tableau
• HTML/CSS
• Javascript
• Bootstrap
• D3.js

Machine Learning Models

We tried following 5 models to predict discrete value (Fatal or Non-Fatal). Our final choice is Decision Tree model.

• Logistic Regression
• SVM
• Neural Network
• Decision Tree
• Random Forest

Flask

Our Flask app hosts our API endpoints as follows:

• /: The root endpoint directs the user to the landing page (index.html) where they are able to provide user input.

• /historicaldata: This endpoint shows the data used in our machine learning models as an HTML table that can be filtered.

• /dashboard: Our Tableau story and dashboards are embedded in the HTML page hosted at this endpoint.

• /models: We link our Jupyter notebooks to all the machine learning models that were tested during the project.

• /predict: This endpoint gets user input from the HTML forms and encodes the values into binary value arrays using a predefined function. These arrays are then merged together to form a 2D array with 66 features as required by the ML model. model.predict gives us our prediction of "Fatal" or "Non-Fatal Injury" which is then passed to the prediction.html page using render_template.