Boston StreetCaster is an ML-based solution to automate the identification of damaged sidewalks in the City of Boston to ensure municipal services are equitably distributed to all residents.
- The Google Street View data collection
- The deep learning models and dataset
- The prototype interface
As the ultimate goal of this project is to identify damaged sidewalk all across Boston we leveraged the City of Boston’s pre-existing geographic sidewalk information to partition sidewalks into GSV images. These images can then be fed into the developed ML models and audited for damage.
See here for more info on this section.
The Machine Learning (ML) Models: Three ML models were trained: two segmentation models and one standard CNN(ResNet) to act as a damage classifier. Data is first fed through a segmentation model to isolate the sidewalk and is then run through the damage classifer to assess any potential issues in the sidewalk.
- See here for more info on the segmentation models.
- See here for more info on the damage classification model.
The Dataset: To build a dataset of damaged sidewalks we contacted a group at the University of Washington that uses Google StreetView (GSV) to find accessibility issues. With their help we were able to collect ~2,000 images of damaged sidewalks. Another ~2,000 images of undamaged sidewalks were randomly sampled from Boston streets to create a balanced dataset. All images were then segmented so any subsequently trained models only learn on sidewalks.
Access the dataset here and find more stats on the dataset.
Our prototype interface allows users to upload images to a server which then segments the sidewalk and classifies into a damaged or not damaged category. This acts as a proof-of-concept for the union of these three parts of the project.
Access more info on the interface here.
Thank you to Project Sidewalk for access the data!
Thank you to the following repos that were modified for our particular ML models. https://github.com/jfzhang95/pytorch-deeplab-xception and https://github.com/fregu856/deeplabv3.