Who gets hurt in Chicago traffic crashes?
Final design, sheet 5 of 5
Oleksandr Kostianets(599443)
Vadym Tovtyn(599435)
Dmytro Levchyk(599437)
Pedestrians are 2.6 % of crashes, but 30 % of the deaths.
of all crashes2.6 %of severe crashes25 %of deaths30 %
Chicago police crash reports, 2019โ€“2025 ยท 752,319 crashes with a location
severe = fatal or incapacitating injury ยท counts, not risk: there is no traffic volume data
Five questions, five views
  1. T1 where(1) map of clusters
  2. T2 when(2) radial, months ร— years
  3. T3 how bad(3) severity split
  4. T4 how(4) ranked scenarios
  5. T5 who(5) age and sex
12345
One screen: filters on top, findings as titles
pedestriancyclistvehicle-onlyCyclistsredraw every view in blue2021fatal + incapacitatingfatal onlyarea: downtown ร—
Road user: click cyclist, every view redraws in blue
Years and severity filter every view
Selections become chips; Reset all clears them
Every card title is the finding, in one sentence
the picked road user, in its colour
the evidence: the chart below shows it
One colour, one meaning
Where and when
area: downtown ร—
Click a cluster: the map zooms in and it splits.
How bad, how, and who
Pedestrian crashes end severe 14ร— as often as vehicle-only ones.
Most common: a car going straight. Most severe: a truck turning right (31 %).
Injured pedestrians: 39 % are aged 20โ€“39; women of 20โ€“29 lead.
Everything is linked
  1. Pick a road user: every view redraws.
  2. Click a cluster: a chip, other views filter.
  3. Click a scenario row: a card opens.
  4. Drag across hours: another chip.
  5. ร— removes a chip, Reset all clears.
area: downtown ร—Truck ยท turning right ร—hours 15โ€“18 ร—
How we will build it
Once, before launch

Data Portal

Chicago, 3 CSV tables: 752,319 crashes, 1,537,399 units, 1,647,604 people

preprocess.py

DuckDB: clean, derive, fold; drop 6,647 crashes without coordinates

3 Parquet files

one road user and one severity rank per crash

On every click
one Docker container

Browser

React 19 + TypeScript; MapLibre GL + deck.gl map; Observable Plot charts

FastAPI

filter state in, one SQL query per view out, cached

DuckDB

reads the Parquet files; map clusters are built in SQL

filters
SQL
rows
5 views
Risk: 719 k vehicle-only crashes, so clustering runs in SQL, not in the browser; groups under 30 crashes are hidden.
Thank you! Questions?