Laser-augmented omnidirectional vision for 3D localisation and mapping

Abstract

The problem of estimating the motion of a robot and simultaneously building a representation of the environment (known as SLAM: Simultaneous Localisation And Mapping) is often considered as an essential topic of research to build fully autonomous systems that do not require any prior knowledge of the environment to fulfill their tasks. The evolution of SLAM is closely linked to the sensors used. Sonars with odometry are often presented as the first sensors having led to convincing results. Since then, 2D laser range finders have often replaced sonars when possible because of the higher precision and better signal to noise ratio. However 2D lasers alone limit SLAM to planar motion estimation and do not provide sufficiently rich information to reliably identify previously explored regions. These observations have led us to explore throughout this thesis how to combine an omnidirectional camera with a laser range finder to help solve some of the challenges of SLAM in large-scale complex environments. The contributions of this thesis concern a method to calibrate central catadioptric cameras (with the development of an opensource toolbox available on the author's website) and find the relative position between an omnidirectional sensor and a laser range finder. How to represent lines and planes for motion estimation is also studied with the use of Lie algebras to provide a minimal parameterisation. Finally we will detail how laser and vision can be combined for planar SLAM and 3D structure from motion.