A System for Automatic Pose-Estimation from a Single Image in a City Scene

摘要

We describe an automatic system for pose-estimation from a single image in a city scene. Each building has a model consisting of a number of parallel planes associated with it. The homographies for the best match of the planes to the image is estimated automatically for each of the possible buildings. We show how the estimation of homographies can be done effectively by reducing the search space and using fast convolution. The model having the best match is then used to determine the position and orientation of the camera. The results of a number of experiments of the system in realistic circumstances is also presented.