Rapid room understanding from wide-angle vision

Abstract

There is an increasing pressure on mobile robotics, especially of the low-cost variety, to perform tasks, with ever increasing complexity, in an uncontrolled environment. Not so long ago mobile robots were expensive devices, employed rarely in the space exploration or industry, where the environment was created to be predictable and the cost was not a major factor. This is not valid anymore, as we see larger and larger numbers of robots reaching the service sector and end consumers. Examples come every year, from vacuum cleaners to the recent advances in autonomous cars.

This thesis focuses precisely on mobile robot sensing capabilities and the ways to extend it. Our particular focus is on a low cost household mobile robots equipped with an omnidirectional camera, enabling extremely wide field of view. As this type of vision sensing is, while not entirely novel, still rather uncommon in the computer vision literature, we had to face multiple challanges due to the lack of reference datasets, algorithm implementations or ground truth sources.

These adversities shaped this thesis to a very large extent, therefore the structure mimicks the progress of computer vision that was done on the classical cameras, so we were able to reach higher levels. We start with a presentation of camera models and calibration techniques. Then we present both sparse and dense SLAM pipelines, allowing us to estimate the poses of the camera accurately and reconstruct dense depth maps respectively. Based on such foundations, we present an occupancy grid free space mapping method followed by a room shape estimation method, both purely relying on omnidirectional vision inputs.

All the presented methods were experimentally verified on synthetic data and on a large number of real world datasets, spanning various sizes and environment types. A separate chapter solely focuses on the engineering efforts required to build the necessary platforms, computing techniques and obtaining valid ground truth.