Calibrated model explorer · 151,968 street segments
Calibrated pedestrian volume estimates for every street segment in Maine, from a Gamma model fitted to StreetLight counts. City Form Lab, MIT · August 2026.
A penalised Gamma generalised linear model with an identity link and log-normally shrunk non-negative coefficients, fitted by maximum a posteriori. Observed volume on a segment is a background constant plus the summed contributions of 56 simulated trip types:
mu = a + SUM_k beta_k × √(flow_k)
Fitted separately for each season on StreetLight all-day pedestrian volumes, validated by spatially blocked ten-fold cross-validation. Out-of-sample Spearman runs 0.72 (summer) to 0.75 (winter).
est_{season}The model's estimate of daily pedestrians on each segment, in StreetLight units.
The background constant is excluded, so these are volumes attributable to simulated trips.
A segment with no simulated trip passing through reads zero — 22% of the network.
Seasonal avg (est_all_year) is the mean of the four seasonal estimates.
c_{season}_{trip type}How much each of the 56 trip types contributes to a segment's estimate, as
coefficient × √(simulated flow). Across a season these sum exactly to the
calibrated estimate in (a). Use this to ask what kind of walking a street carries — commuting,
errands, waterfront leisure — rather than just how much.
Coefficients are not comparable across trip types in absolute terms: each simulated flow is on its own scale. Within a single segment, the contributions are directly comparable.
{season}_camera_adjStreetLight is itself a modelled index, not a headcount. Comparing 190 camera sites against the calibrated estimates gives a conversion to observed-count units:
camera equivalent = 0.0205 × estimate1.40
Use with care. The conversion was fitted on cameras whose estimates ran 1.25–7,920, median about 950. The statewide median estimate is 3.3, so on most of the network this extrapolates well below its fitted range and reads near zero. It also does not remove the model's compression: quiet streets still read high and busy streets low.
y_{season}The observed values the model was fitted against — all-day pedestrian volume averaged across measurement sites. These are measurements, not estimates, and exist on only 10,906 segments (7.2% of the network). Everything else is blank. Useful for checking the model against truth where truth exists.
178 intersections across Maine where pedestrians were counted by camera. Unlike everything
else on this map these are observed people, not estimates and not a mobile-device index —
the most authoritative measurement available here. Circle size and colour both show
Avg_FullDay, the average daily count, which runs from 10 to 8,865 (median 156).
Hover or click a camera for the season it was captured in, its lunchtime and evening counts,
and how many hours were observed.
Season comes from the observation dates: 94 summer, 47 autumn, 28 spring, 2 winter, and 3 that span two seasons; 4 cameras could not be dated. Counts are not seasonally adjusted — each is what that camera saw at that time of year, so compare a camera against the matching seasonal estimate rather than against the seasonal average.
Two cautions. Observation length varies from 7 to 24 hours, and the full-day figure is scaled from whatever window was watched, so short observations carry more uncertainty. And these counts are what the camera conversion in (c) was fitted against — so comparing them against the camera-adjusted layer is not an independent test.