Crowd Counting through Remote Sensing

With global development, more and more people group up around large cities and go to major events such as sports competitions (Olympics, World Cup), festivals, concerts, demonstrations and political, cultural, social or religious gatherings. It then becomes increasingly important to be able to monitor, count and analyse such large crowds, for example to avoid over-crowding (panic in large crowds has sometimes resulted in several casualties) or to report exact numbers in the media (often used for so-cial/political propaganda).

Currently, crowd counting is performed on single-view images. This process is both limited by a large annotation requirement (all heads in the images need to be manually labelled with a dot), but also by not being able to fully capture the entire crowd in one image. To solve the latter limitation, researchers have also investigated multi-view crowd counting, where the whole crowd is seen from various view-points. However, these approaches require datasets with even more annotations, which becomes in-creasingly more complex when annotating from multiple views. In simpler terms, we are far from be-ing able to estimate the full size of a crowd at a large gathering, such as the crowning of HM King Frederik X. One can only come up with a broad estimate.

The goal of this PhD project is to develop novel methods for crowd counting when large crowds are involved, and where varying resolution makes it extremely difficult to apply current SOTA models. As part of the project, we also want to push research in remote sensing, by expanding analysis and counting to objects seen from satellites. In the process, we also aim to explore whether counting peo-ple from high resolution satellite images is a plausible method.