Communications of the ACM
AI Could Transform the Science of Counting Crowds
New software can help scientists count crowds.
Credit: Noam Galai/WireImage/Getty Images
The University of Central Florida (UCF) has been developing artificial intelligence (AI) solutions to relieve crowd scientists from the task of manually counting heads.
Machine-learning software can provide automated headcount estimates for large crowds based on aerial images.
Like the manual approach, the UCF software subdivides a given crowd image into smaller patches, and the individual patch counts are then averaged together based on assumptions about crowd density.
To boost the system's accuracy, researchers have switched to a deep-learning AI that takes advantage of neural networks to process images. However, this method still presents some of the same challenges human crowd scientists face.
Ideal crowd images are taken from above by drone, aircraft, or satellite, which may not always be possible for events taking place in restricted air zones. Meanwhile, images taken from a side angle may distort a computer's sense of perspective and scale.
One of the biggest challenges for the AI approach to automated crowd counting is the need for large amounts of training data that have been accurately annotated by hand, and the UCF team plans to use online crowdsourcing services to help manually create those datasets.
From IEEE Spectrum
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found