Use of Scene Geometry Priors for Data Association in Egocentric Views

The joint use of dynamic, egocentric view cameras and of traditional overview surveillance cameras in high-risk contexts has become a promising avenue for advancing public safety and security applications, as it provides more accurate localization and finer analysis of individual interactions. However, the strong scene scale changes, occlusions and appearance variations make the egocentric data association more difficult than the standard across-views data association. To address this issue, we propose to use two independent geometric priors and integrate them with the classic appearance cues into the objection function of data association algorithm. Our results show that the proposed method achieves significant improvement in terms of the association accuracy. We highlight the attractive use of geometric priors in across-views data association and its potential for supporting pedestrian tracking in this context.