Robust camera pose recovery using stochastic geometry
Matthew Antone, Seth J. Teller · 2001
The Problem: The objective of 3-D machine vision is to infer geometric properties (e.g. shape and size) and photometric attributes (e.g. color, texture, reflectance) from a set of 2-D images. Every such vision task relies on accurate camera calibration, that is, knowledge of the camera’s intrinsic parameters (focal length, lens distortion, etc.) and extrinsic parameters—orientation, position, and scale relative to a fixed frame of reference. This research concerns the automatic recovery of precise extrinsic pose among a large set of images, assuming that accurate intrinsic parameters and rough estimates of extrinsic parameters are available. This work also investigates models of geometric uncertainty and the application of projective inference techniques. Motivation: The MIT City Project [2] employs a set of unique algorithms to produce accurate, metric reconstruction of urban landscapes from pose imagery, or images annotated with approximate geo-referenced position and orientation. All components of the system are completely automatic except for extrinsic camera registration, which is obtained via manual correspondence of point features across images followed by global bundle adjustment. The acquired data consists of a large set (tens of thousands) of images, making the the human-assisted component quite tedious. Thus the primary goal of this work is to provide fully automatic, scalable external registration. Previous Work: There exists an enormous body of work in automatic camera registration, which is most often coupled with the recovery of 3-D structure. Some techniques use known calibration targets to determine camera pose [1].