Scalable, Absolute Position Recovery for Omni-Directional Image Networks

We describe a linear-time algorithm that recovers absolute camera positions for networks of thousands of terrestrial images spanning hundreds of meters, in outdoor urban scenes, under varying lighting conditions. The algorithm requires no human input or interaction. It is robust to up to 80% outliers for synthetic data. For real data, it recovers camera pose which is globally consistent on average to roughly 0.1 and five centimeters, or about four pixels of epipolar alignment, expending a few CPUhours of computation on a 250MHz processor. This paper’s principal contributions include an extension of Monte Carlo Markov Chain estimation techniques to the case of unknown numbers of feature points, unknown occlusion and deocclusion, and large scale (thousands of images, and hundreds of thousands of point features) and dimensional extent (tens of meters of inter-camera baseline, and hundreds of meters of baseline overall). Also, a principled method is given to manage uncertainty on the sphere of directions; a new use of the Hough Transform is proposed; and a method for aggregating local baseline constraints into a globally consistent constraint set is described. The algorithm takes intrinsic calibration information, and a connected, rotationally registered image network as input. It then assembles local, purely translational motion estimates into a global constraint set, and determines camera positions with respect to a single scene-wide coordinate system. The algorithm’s output is an assignment of metric, accurate 6-DOF camera pose, along with its uncertainty, to every image. We assume that the scene exhibits local point features for probabilistic matching, and that adjacent cameras observe overlapping portions of the scene; no further assumptions are made about scene structure, illumination conditions, or camera motion. We assess the algorithm’s performance on synthetic and real data, and demonstrate several results. First, wide-FOV imagery makes registration fundamentally more robust against failure, and more accurate, than ordinary imagery. Second, we show that by combining thousands of noisy, gradient-based (point) features into a small number of projective motion estimates (baselines), the algorithm achieves accurate registration even in the face of significant lighting variations, low-level feature noise, and errors in initial position estimates.