Pedestrian Activity & San Francisco's Hills
Do San Francisco's steep blocks (≥20% grade) keep people from walking through them? We expected the hills to act as walls. The data pushed back: nearly a third of recorded trips crossed 'significant' slopes, and half of running trips did.
Research Question
San Francisco has hills - specifically, blocks with slopes of 20% or greater, which the city's planning department calls "significant" and regulates accordingly. We wanted to know how that terrain correlates with where people actually walk and run, using the AOMA cell-phone localization sample.
The test set: 3,000 recorded trips, cleaned in Python and joined against the city's official slope shapefiles. Visualized in Kepler.gl.
Data Cleaning
The raw dataset included trips that were, at times, comically superhuman. We validated every trip against human-probable movement: no segment faster than 10.44 meters per second (Usain Bolt's world-record average over 100 meters) and no stride longer than 2.45 meters, then clipped everything to the city boundary with a 500-meter accuracy buffer. 3,000 files in, 2,009 probable trips out - 1,880 walking, 129 running.
What the Hills Actually Do
- The hills don't empty out. About 31% of recorded trips passed through areas at or above 20% grade, and roughly half of the total distance traveled in the sample (50.6%) crossed high-slope terrain. For a city that regulates these blocks as exceptional, people move through them constantly.
- Runners seek out what walkers tolerate. About 30% of walking trips overlapped high-slope areas - but 50% of running trips did. The steepest terrain in the city doubles as its training ground.
- Crossing the hills costs something measurable. Trips passing through high-slope areas ran longer on every dimension we could measure - more calories, more distance, more elapsed time - than the sample as a whole.
Limitations
- The validation caps are blunt. World-record speed and stride are conservative, rudimentary ceilings - and debatably exclusionary of non-normative bodies. A probability distribution over speed and stride would be a more equitable filter than hard caps.
- The sample self-selected. AOMA users opted into sensor tracking. Nothing here generalizes to all pedestrians - it describes the people who chose to be recorded.
- The location data is fuzzy by design. The dataset strips 0-100 meters from each trip's start and end for anonymization, and our 500-meter accuracy buffer stacks its own assumption on top. We also standardized start times, trading away temporal analysis for a cleaner spatial one.
