Towards View Invariant Person Counting and Crowd Density Estimation for Remote Vision-Based Services

Crowd monitoring in mass events is a highly important technology to support the safety of event attending persons. Proposed methods are often limited to one specific viewing condition and have to be retrained or even redesigned if the viewing angle is changing which is particularly mandatory in airborne based applications. We present a novel framework for highly view invariant person counting and crowd density estimation from single airborne or terrestrial images based on a generalized human head detector and a regression based density estimate. Employing manually labeled reference data, we present detailed accuracy analyses for object detection and for density based person counting. The resulting human counter demonstrates a mean error of 5% over three different data sets. At the same time it thus provides a highly efficient quality indicator for benchmarking security critical decision support services.