Road safety

The purpose of this project is to use regression modeling and neural networks to predict collision rates in New York City using public road data and other geological features, and to identify key road infrastructures that can improve safety as economically as possible.

Authors

• Assaf Bar-Natan
• Jesse Frohlich
• Jacob van Hook
• Pedro Lemos

Description

Overview

Traffic collisions are a leading cause of death and injury amongst Americans under the age of 54. A significant portion of these collisions are preventable by changing driver attitudes and improving in-vehicle safety measures, but also by creating better public infrastructure that encourages safe practices.

The goal of this project is to develop a predictive model that will assess the change in collision rates in a given road section of New York City after changes in road features are implemented.

We will use data from NYC Open Data and apply machine learning techniques in order to model collision rates on roads.

Stakeholders: General public, local politicians, neighbourhood advocacy groups, commuters, urban planners, road design engineers, NYC police, insurance companies.

KPIs for collision rate forecasting: mean squared error (MSE) and mean absolute error (MAE) between the true collision rate and predicted collision rate.

Datasets

Our data for this project is coming from NYC Open Data. Their data includes features such as:

• Road Width
• Traffic Volumes
• Existence of Speed Humps
• Existence of Bike Lanes
• Existence of Trees
• Vehicle collision locations and times

The dataset used to train and test our models is built by joining and aggregating information contained in different data sources. This is done automatically by src/data_generator.py. This script will be executed by running make in the top directory.

Approach

We have several proposed models we are going to test to see how well they predict the rate of collisions on a given road segment.

• Baseline Model: This model always predicts the mean collision rate across the entire city for every road segment, regardless of features. This is important to have because it gives us a basis of comparison by which to measure our other models. If a model cannot do better than just predicting the average every time, it is likely not a very good model.
• Linear Regression: This straightforward model uses linear regression and predicts a linear relationship between features, or combinations thereof, and the collision rate. This is also a good basis of comparison and can be used to assess the effectiveness of more complicated models.
• Random Forest Regression and XGBoost: We also want to test the effectiveness of a random forest regressor. This may prove useful as a way to determine relationships which we would not be able to see through standard linear regressions.
• K-Nearest Neighbors Regression: We want to look at the predictive power of a K-Nearest Neighbors model because roads with similar features are likely to have similar collision rates and can therefore serve as a predictor.

Future Iterations

• Adding more features relevant to road safety, for instance:
  • the installation of leading pedestrian traffic signals
  • the amount of physical separation present in cycling infrastructure
  • the width of individual lanes, as opposed to the entire roadway
• Extending the model to other cities with "Vision Zero" road safety plans (and hence sufficient data to analyse).

Project outline

Analyse features of a road's design, and build a model to predict the safety of the road, either using KSI metrics or average/ maximum vehicle speed as proxies for safety & comfort.

Usage

To set up the working environment, it is strongly encouraged to set up a virtual environment, say with pyenv:

$ python -m venv .venv

To activate the virtual environment, source the following:

$ source .venv/bin/activate

Finally, you can install the packages locally via pip:

pip install -r requirements.txt

The OpenDataDownloader class in src takes an optional app token that can be obtained from the NYC Open Data website. To run some of the notebooks, you will need to store this app token in a file named .env in the top directory with the following content:

NYC_OPENDATA_APPTOKEN=<insert your app token here>

Finally, we have a make file to download all the data and aggregate it into a cleaned dataset, you need only run make from the root project directory.

Contributing

This project uses pre-commit to ensure code formatting is consistent before committing. The currently-used hooks are:

• black for *.py or *.ipynb files.
• prettier (with --prose-wrap always for markdown)

To set up the hooks on your local machine, install pre-commit, then run pre-commit install to install the formatters that will run before each commit.