USP counts Brazil's protest crowds with drone AI — and cut inflated estimates to size

In late February 2024, throngs filled São Paulo's Paulista Avenue in support of former president Jair Bolsonaro. His team declared it 'the largest political event ever held' on the avenue; the São Paulo state public security department, governed by Bolsonaro allies, claimed 600,000 attendees. A team at the University of São Paulo put the real figure at 185,000 — less than a third — using a tool that identifies individuals in drone images and counts them with AI.

The tool was built by political scientist Pablo Ortellado's team on a Chinese counting framework that had previously been used to count heads of cattle — local critics call Bolsonaro's supporters gado, Portuguese for cattle. The standard method, density times occupied area, 'is usually a guess', Ortellado said; his team's count is 'much more accurate'. The team has attended far-right and left-wing demonstrations since 2015 to study polarization, frustrated that official and media rally figures were highly discrepant — crucial information for understanding politics.

They launched the tool in 2022 and have updated it since, reducing the margin of error from 30% to 12%. The stakes are global: protests more than tripled between 2006 and 2020 across 101 countries, and crowd estimates routinely diverge by tens of thousands, sometimes millions, shaping perceptions among voters and analysts.

The method is spreading: analysts and newsrooms in Brazil have started replicating it and citing its estimates in reports and papers. 'Before the tool, these figures were always tainted by political interests, both from the government and the protests' organizers,' said researcher Jonas Medeiros. The São Paulo security department said it uses ground agents plus drone and helicopter imagery with estimation software, but did not answer questions about the gap with USP's count; Bolsonaro's lawyer alleged the tool undercounts his rallies.

Strap a cattle-counting AI to a drone: it identifies individuals in aerial images and counts them, turning protest sizes from guesses into measurements

The pivot is the trick: a framework built to count cattle now counts protesters, and no one had applied it to demonstrations.

Detection beats estimation: identifying individuals in aerial images removes the density-assumption step where politics creeps in.

The error is quantified: a stated margin — cut from 30% to 12% — makes the number arguable on evidence rather than assertion.

It works on both sides: the team applies it to left-wing and far-right rallies alike, which protects its credibility.

When a contested number is a political weapon, move the measurement from estimation to detection: counting heads directly is harder to argue with than multiplying guesses.

The team keeps updating the tool; analysts and newsrooms began replicating the method and citing its counts in reports and papers. The São Paulo public security department did not respond to questions about the discrepancy with USP's estimate, and Fabio Wajngarten, Bolsonaro's lawyer, alleged the tool underestimates attendance without answering Rest of World's request for comment.