In recent weeks, a method of counting people at events used by the University of São Paulo (USP) gained repercussion, after estimating a more realistic number – 185 thousand people – than the São Paulo Public Security Secretariat – 600 thousand people –, during a political event held on Avenida Paulista on February 25th.
One of the highlights reported about the tool is the use of Artificial Intelligence to carry out the accounting. After all, how does it work and what technologies does it use to do this? Detective TC has separated the main information for you below.
What is the project?
The project that gave rise to the estimate is the Political Debate Monitor. It was created by USP's School of Arts, Sciences and Humanities (EACH), coordinated by Pablo Ortellado and Márcio Moretto.
The initiative has been operating since April 2015, in an event that took place that month, in which field research was carried out with the individuals who participated in the demonstration.
However, the estimates only began in a demonstration that took place in August 2022, in the external area of the USP Law School, in Largo de São Francisco. Since then, the project has used the technique eight times to count people at events of this type.
In an interview with Estadão, Pablo Ortellado reinforced the importance of the project, as a way of applying a method for counting events, which ends up not supporting the grandiose numbers that are usually released.
“A political culture of grandiose numbers was created, not only in political demonstrations but on other occasions, such as Virada Cultural or carnival blocks. These are very big numbers that don’t hold up when you rely on some criteria.”
Paul Ortellado
Professor of Public Policy Management and study coordinator at USP
How does it work in practice?
The count is done using high-resolution photos taken over the desired area – in the most recent case, from the end of February, 43 images were recorded, of which 11 covered the entire area. Records take place using drones in a vertical position, divided by perimeter.
With the material done, without any overlap, the captures are sectorized and applied using the method called Point to Point Network. It consists of software that detects the presence of heads and estimates the number of people.
The Chinese program was initially launched in 2021, during the International Conference on Computer Vision, a conference held in the city of Montreal, Canada.
Artificial intelligence
One of the highlights of Point to Point Network is the use of Artificial Intelligence to make estimates. The tool takes advantage of the most advanced technology when processing data.
To achieve this, the software underwent manual training, focusing on recognizing heads. However, there are still its difficulties. The program, for example, cannot identify individuals located under trees or other structures. There is also doubt regarding the recognition of people wearing caps.
Tech
27 Fev
Economy and market 15 Feb
Precision and accuracy
Despite the difficulties that still exist and can be improved for the future, Point to Point Network is considered software with “the best precision and accuracy values for identifying heads in crowd photos”, as researcher Márcio Moretto highlighted to Poder360.
The method used by USP has an accuracy rate of 72.9% and an accuracy – level of accuracy – of 69.5% in detecting each individual, as experts highlight.
When counting the public, the average absolute percentage error consists of 12% plus or minus, in the case of aerial images with more than 500 people.
What is your assessment of the software used by USP to count people in acts? Tell us!