Establishing Correspondence in distributed cameras by observing humans

Correspondence between distributed cameras with overlapping views is needed for several tasks in surveillance and smart environments. This paper proposes an adaptive correspondence estimation technique by observing multiple humans in a planar scene. In contrast to other RANSAC-based approaches, it does not require prior information about corresponding features. The proposed techniques uses only results from a person detector and scene specific detection filtering for estimating the inter-image homography. The method is self-configurable, adaptive and provides robustness over time by exploiting temporal and geometric constraints. The correspondence is accurately estimated in spite of error sources such as missed detections, false detections and non overlapping fields of view. Results on a variety of datasets demonstrate the general applicability. Experiments show that the proposed correspondence estimation approach outperforms a common baseline approach.