Main idea: Based on the computation of the vanishing points (VP) and their matching through different images coming from different cameras, (hemispherically-tiled set of images captured from a single position in space) the computation of the relative camera rotation is performed.
This work first compute the vanishing points present in each image based on the direction of the edges extracted from the scene. The representation used for these edges is the intersection of the plane through the edge and the focal point with the Gaussian sphere. The direction of the 3D edges (vanishing points) is computed using an EM approach.
The vanishing point estimation is composed of two tightly-coupled sub-problems: classification (grouping observed edges into parallel sets) and estimation (finding the best VP for each set). The authors present an hybrid approach to VP detection and estimation which combines the Hough transform (HT) for detection with the acccuracy of least squares for estimation. An EM algorithm is formulated to probabilistically model edges and their directional uncertainty. A final verification step rejects spurious directions.
link: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=854809
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment