Crosswalk detection and accessible routing

Context

For nearly three years at Georgia Tech, I worked as an undergraduate researcher in computer vision and urban mobility, on a multidisciplinary team studying pedestrian infrastructure. The question behind the work is simple to ask and hard to answer at city scale: where can a person who uses a wheelchair actually go?

Answering it needs to know where the crossings are, whether the routes between them are usable, and what it would cost to fix the gaps. Most of that information is missing, out of date, or locked in imagery that no one has the time to read by hand.

What I did

The work ran across three connected threads.

1. Detecting crosswalks in aerial imagery. The goal was to find crosswalks automatically in aerial images, which are often noisy and low resolution: markings are faded, partly hidden by trees, cars and shadows, or only a few pixels wide. I worked on a supervised learning pipeline for this, and the largest gain came from the data rather than the model: by rebuilding the training dataset and refining the pipeline around it, detection accuracy rose from a 25% baseline to 60%.

2. Accessible navigation. Working with the team, I helped enrich OpenStreetMap data and apply graph algorithms to find routes a wheelchair user could actually take, using Gwinnett County, Georgia, as the study area. Street maps are built for cars; a route that is short on the map may be impossible in practice if a crossing or curb ramp is missing.

3. Costing pedestrian infrastructure. I carried out field data collection in person across several cities, and built tools that estimate the equipment, labour and travel cost of surveying and improving pedestrian infrastructure.

Across all three, I built reproducible reporting and visualisation workflows that the wider research team used, so results could be regenerated and compared rather than rebuilt by hand.

Key decisions and tradeoffs

What I took from it

• Data before models. The biggest improvement in detection came from rebuilding the dataset, not from a cleverer architecture. When the input is noisy, the training data is where the leverage is.
• Maps are not the street. Map data and the ground often disagree, and the disagreements are exactly what matter to someone who cannot step off a curb. Field collection is slow, but it is the only way to know.
• Make results reproducible. In a team spanning several disciplines, a result that only one person can regenerate is a result nobody can build on.

This work is also where the idea for this site came from: the hero is a city built from real map data, which is where this research started.

Results

• Crosswalk detection accuracy on noisy, low resolution aerial imagery raised from a 25% baseline to 60%.
• Accessible routing explored on real OpenStreetMap data for Gwinnett County.
• Cost estimation tools for pedestrian infrastructure surveys.
• Reporting and visualisation workflows adopted by a multidisciplinary team.