Mapping large environments with an omnivideo camera

We study the problem of mapping a large indoor environment using an omnivideo camera. Local features from omnivideo images and epipolar geometry are used to compute the relative pose between pairs of images. These poses are then used in an Extended Information Filter using a trajectory based representation where only the robot poses corresponding to captured images are reconstructed. The features with the geometric constraints also give a robust similarity measure that is used for data association. Our experiments show that an accurate map can be built in real time of a small office environment. For large environments, big loops can be closed and a map can be built in nearly linear time.