Skip to content
Vincent Wren
← All work
Python May 2026 – Present

Vision Zero Houston

Pharis Fellowship research asking whether Houston knows where its dangerous streets are.

  • Python
  • Computer Vision
  • Statistics
  • Transportation Safety
  • Research
Two maps of Houston side by side: the City’s crash-based High Injury Network and this project’s design-based Proactive Safety Network, with held-out capture rates of 46% and 51%.

For my Pharis Fellowship with the University of Houston Honors College and HPE Data Science Institute, in partnership with the office of Council Member Joseph Panzarella (District C), I built a citywide answer to one question: does Houston know where its dangerous streets are? The project has two halves. The first is Vision Zero Houston, a public dashboard that makes 421,679 state crash records (2016 to mid-2026) across 66,922 street segments explorable by street, council district, neighborhood, travel mode, and time of day, with one-click printable reports for council use. The second is a proactive street-design risk model: a negative binomial crash-frequency regression, supplemented with computer-vision features (SegFormer and Faster R-CNN) extracted from 59,261 street images, validated spatially on held-out council districts and temporally on crashes the model never saw. Frozen at the end of 2021 and graded on 2023 to mid-2026, the design model captured 51% of future severe crashes versus 46% for the City’s crash-history High Injury Network at matched mileage, and flagged roughly 300 miles of high-risk streets missing from the official network, where 574 severe crashes later occurred. Street design, not crash history alone, tells a city where harm lands next. I sole-authored the resulting paper, submitted to the Transportation Research Board (TRB) 2027 Annual Meeting.

Highlights

  • Built a public citywide dashboard making 421,679 crash records across 66,922 street segments explorable by street, district, neighborhood, mode, and time.
  • Modeled crash risk from street design: negative binomial regression plus computer-vision features (SegFormer, Faster R-CNN) from 59,261 street images.
  • Validated spatially and temporally: the design model captured 51% of held-out severe crashes vs 46% for the City’s High Injury Network.
  • Identified roughly 300 miles of high-risk streets missing from the official network, where 574 severe crashes later occurred.
  • Sole-authored the paper, submitted to the TRB 2027 Annual Meeting.

Open the live dashboard ↗ Read the paper (PDF) ↗ View the code on GitHub ↗