Automatic Recovery of Camera Positions in Urban Scenes

Abstract

Accurate camera calibration is crucial to the reconstruction of three-dimensional geometry and the recovery of photometric scene properties. Calibration involves the determination of intrinsic parameters (e.g. focal length, principal point, and radial lens distortion) and extrinsic parameters (orientation and position). In urban scenes and other environments containing sufficient geometric structure, it is possible to decouple extrinsic calibration into rotational and translational components that can be treated separately, simplifying the registration problem. Here we present such a decoupled formulation and describe methods for automatically recovering the positions of a large set of cameras given intrinsic calibration, relative rotations, and approximate positions. Our algorithm first estimates the directions of translation (up to an unknown scale factor) between adjacent camera pairs using point features but without requiring explicit correspondence between them. This technique combines the robustness and simplicity of a Hough transform with the accuracy of Monte Carlo expectation maximization. We then find a set of distances between the pairs that produces globally-consistent camera positions. Novel uncertainty formulations and match plausibility criteria improve reliability and accuracy. We assess our system's performance using both synthetic data and a large set of real panoramic imagery. The system produces camera positions accurate to within 5 centimeters in image networks extending over hundreds of meters.