Late fusion for person detection in camera networks

In this paper, we present a novel method to detect multiple partially occluded persons in multi-view camera networks. We present a new fusion scheme to integrate the output of part-based object detectors from multiple camera views. This is achieved using subtle and precise modeling of detection and projection uncertainties as well as a fusion method based on probabilistic kernel density estimation. Using a multi-view setup also allows to incorporate additional real-world prior knowledge about person appearances, which not only speeds up processing, but also increases detection rates. Experiments show that this multi-camera approach outperforms methods based on a single perspective, particularly in occlusion-intense scenarios.