Reckless motion estimation from omnidirectional image and inertial measurements

Workshop on Omnidirectional Vision and Camera Networks

物理学;材料科学

Dennis Strelow ; Sanjiv Singh

PID  :  83315

学科分类：材料科学（综合）

来源: CEUR

【 摘 要 】

Two approaches to improving the accuracy of camera motion estimation from image sequences are the use of om nidirectional cameras, which combine a conventional cam era with a convex mirror that magnifies the field of view, and the use of both image and inertial measurements, which are highly complementary. In this paper, we describe optimal batch algorithms for estimating motion and scene structure from either conventional or omnidirectional images, with or without inertial data. We also present a method for motion estimation from in ertial data and the tangential components of image projec tions. Tangential components are identical across a wide range of conventional and omnidirectional projection mod els, so the resulting method does not require any accurate projection model. Because this method discards half of the projection data (i.e., the radial components) and can oper ate with a projection model that may grossly mismodel the actual camera behavior, we call the method “reckless” mo tion estimation, but we show that the camera positions and scene structure estimated using this method can be quite