Applying information theory to efficient SLAM

Abstract

The problem of autonomous navigation of a mobile device is at the heart of the more general issue of spatial awareness and is now a well-studied problem in the robotics community. Following a plethora of approaches throughout the history of this research, recently, implementations have been converging towards vision-based methods. While the primary reason for this success is the enormous amount of information content encrypted in images, this is also the main obstacle in achieving faster and better solutions. The growing demand for high-performance systems able to run on affordable hardware pushes algorithms to the limits, imposing the need for more effective approximations within the estimation process. The biggest challenge lies in achieving a balance between two competing goals: the optimisation of time complexity and the preservation of the desired precision levels. The key is in agile manipulation of data, which is the main idea explored in this thesis. Exploiting the power of probabilistic priors in sequential tracking, we conduct a theoretical investigation of the information encoded in measurements and estimates, which provides a deep understanding of the map structure as perceived through the camera lens. Employing information theoretic principles to guide the decisions made throughout the estimation process we demonstrate how this methodology can boost both the efficiency and consistency of algorithms. Focusing on the most challenging processes in a state of the art system, we apply our information theoretic framework to local motion estimation and maintenance of large probabilistic maps. Our investigation gives rise to dynamic algorithms for quality map-partitioning and robust feature mapping in the presence of significant ambiguity and variable camera dynamics. The latter is further explored to achieve scalable performance allowing dense feature matching based on concrete probabilistic decisions.