Automatic Inter-image Homography Estimation from Person Detections

Inter-image homographies are essential for many different tasks involving projective geometry. This paper proposes an adaptive correspondence estimation approach between person detections in a planar scene not relying on correspondence features as it is the case in many other RANSAC-based approaches. The result is a planar inter-image homography calculated from estimated point correspondences. The approach is self-configurable, adaptive and provides robustness over time by exploiting temporal and geometric information. We demonstrate the manifold applicability of the proposed approach on a variety of datasets. Improved results compared to a common baseline approach are shown and the influence of error sources such as missed detections, false detections and non overlapping field of views is investigated.