6533b835fe1ef96bd129f16f

RESEARCH PRODUCT

Calcul de pose d'un réseau de caméras hybrides à partir de droites

Sang Ly Cédric Demonceaux Pascal Vasseur Claude Pegard

subject

[INFO.INFO-CV] Computer Science [cs]/Computer Vision and Pattern Recognition [cs.CV][INFO.INFO-CV]Computer Science [cs]/Computer Vision and Pattern Recognition [cs.CV][ INFO.INFO-CV ] Computer Science [cs]/Computer Vision and Pattern Recognition [cs.CV]

description

We present a pose estimation approach for a setup of multiple hybrid cameras using line features. From line images across at least three single viewpoint cameras such as perspective, central catadioptric or fish-eye ones, the camera poses are recovered by a decoupling of rotations and translations : rotations are estimated from vanishing points of parallel lines and translations are calculated from known rotations and line images. The 3D lines are reconstructed by the triangulation of projective planes passing through line correspondences. The camera poses and 3D scene are finally refined by bundle adjustment. This technique can be used to recover the topology of a multiple hybrid camera system or extended to vision-based robot localization. The main contributions of this approach are the sole use of lines for the estimation, rather than point features and that it can handle different camera types in a unified manner. Moreover, we present an efficient bundle adjustment on camera and 3D line parameters by minimising the error of the line normals on spherical images. The proposed algorithm has been validated on simulated data and real images.

https://hal.inria.fr/inria-00595485/document