Motion Estimation for Hybrid Cameras Using Point and Line Feature Fusion

We present in this paper a structure-from-motion approach for single viewpoint cameras using point and line features. From the image sequence captured by any central projection cameras, the motion can be recovered by decoupling of orientation and displacement: rotation is estimated from vanishing points of parallel lines and translation is calculated from known rotation and point/line correspondences. Next, the 3D structure is reconstructed by triangulation of projective rays or planes. Lastly, the camera motion and 3D scene are optionally refined by bundle adjustment. This approach can be applied to recover the motion of autonomous robots or the arrangement of a vision-based surveillance system equipped with any single viewpoint cameras such as perspective, fish-eye and catadioptric cameras. Moreover, the translation can be estimation from points and/or lines depending on the availability of these features in the images. The proposed algorithm has been validated on simulation data and real images.