Russian friends of mine recently pointed me to a very interesting computer vision program called “White Counter”. The purpose of this mysterious-sounding algorithm was to automatically count people in a dense crowd from video footage. The program developed in early 2012 by two experts in computer vision and artificial intelligence: Anatoliy Katz and Igor Khuraskin. I recently spoke with Anatoliy to learn more about his work given potential applications for counting refugees and internally displaced peoples using UAVs.
He and Igor created their “White Counter” in order to counter government figures on the number of protestors who join demonstrations. As is typical, the Kremlin will always downplay the numbers. So Anatoliy and Igor used insights from fluid dynamics to create their algorithm, measuring average speed of flow as well as density, for example. Note that “White Counter” is not a fully automated solution. The algorithm requires manual counts every 30 seconds in order to estimate overall crowd figures. But the results of the algorithm are impressive: the error margin at this point is less than 3%. Anatoliy and Igo used their algorithm during the “March of Millions” on September 15, 2012 (video above). Their code is python based and open-source, so if you’re interested in experimenting with the code, simply email firstname.lastname@example.org.
My colleague Austin Choi-Fitzpatrick and his team are also working on a similar challenge. They too are interested in estimating the size of social movements. As Austin rightly notes, “Establishing the size of a protest event is critical for social movements as they signal their legitimacy to the media, to the general public & as they demonstrate their strength to the authorities that they’re challenging.” But the methods we use to estimate how large a protest is haven’t changed in more than half-a-century. So Austin & team are looking to update these methods using UAVs and aerial imagery. They take a given image, identify the total area covered by protestors, then slice up the image into a grid of micro-images. They then assess the density level of the crowd in each micro-image. The video below introduces the project and research in more detail.
Perhaps in the near future humanitarian UAVs will be able to draw on these advances in computer vision to assess refugee populations and the number of displaced peoples following major disasters. In the meantime, we can use simple crowdsourcing solutions like MicroMappers to estimate populations. There is a precedence for applying crowdsourcing to compute population counts—see this UN Refugee Project, for example. But if you know of any related work other than the JRC’s efforts that draws on automated techniques, then please let me know, thank you!