The Helmholtz Young Investigators Group “ARIADNE - Aerial Imagery Analytics by Deep Neural Networks” aims at improving mobility and public safety via automated aerial monitoring of urban traffic and crowded public spaces. To this end, we develop novel computer vision algorithms relying on powerful tools from deep learning for the aerial visual understanding of urban traffic and infrastructure, as well as of mass events.
Conventional image processing methods adopted in remote sensing struggle to keep pace with the increased quantity and quality of visual data provided by modern air- and spaceborne sensors. Deep neural networks, responsible for the recent breakthroughs in computer vision, can cope with and even require such large amounts of visual data and hold the promise of bridging the semantic gap for aerial image understanding as well. The research group therefore develops innovative computer vision algorithms to tackle problems such as the exact localization and tracking of traffic participants (vehicles, pedestrians, bicyclists, etc.), the fine-grained segmentation and mapping of road infrastructure, as well as crowd counting and the classification of human crowd behavior at mass events. The developed methods can be applied to acute societal problems such as emergency monitoring, ensuring safety at large public events, and traffic flow and safety monitoring for sustainable mobility
Left: Aerial view of the Bauma 2016 construction trade fair in Munich, Germany, with collected person annotations (marked in red). Right: For a given image patch of a crowd (top, yellow frame in aerial image), we show the ground-truth crowd density map with person locations (middle), as well as the predicted density map (bottom) with estimated person detections