Home / Pedestrian Safety
03 · Safety Modeling

Whose Safety Perception Gets Lost When You Scale It Up?

Vision Zero Complete Streets Active Transportation Planning Equity Analysis

Presented at the ACSP Annual Conference and the AAG Annual Meeting

1

Background & motivation

Complete Streets and Vision Zero prioritization depends on perception data, and community surveys can't supply it at scale anymore — they're too slow and too expensive. So I tested a faster substitute against 477 real Columbus residents (work I later presented at ACSP and AAG). It holds up reasonably well citywide, but it flattens the most dangerous streets toward the middle, and it's least accurate for older, lower-income, and female residents. That's not a minor demographic footnote — it's the exact group Vision Zero is supposed to protect.

2

Screening a whole city

Rating distribution — human vs. five models
Stacked bar chart of safety rating distribution, human survey vs five simulated models
Human survey vs. five models, side by side
Grouped bar charts comparing each model's rating distribution against the human baseline
Same comparison, model by model, against the human baseline (gray)
Spatial pattern across Columbus
Kriged interpolation maps of perceived safety across Columbus, human ratings versus simulated ratings
Interpolated safety surface — human (left) vs. simulated (right)
3

Where it breaks for planning decisions

Where the match falls apart
Line chart of exact-match and within-one accuracy by human-rated safety value across five models
How often the simulated rating matches — by how dangerous residents actually rated the street
Sankey diagram showing where each human safety rating ends up when simulated, pooled across five models
Where each human rating lands when simulated — red overestimates safety, blue underestimates it
Map of Columbus showing the spatial distribution of bias magnitude between simulated and human safety ratings
Where the error concentrates across Columbus — darker means AI and human ratings diverge more