Localization and tracking of parameterized objects in point clouds

Abstract

This thesis focuses on object recognition and tracking from three dimensional point cloud renderings of dense range and bearing data. Sensors like laser range-finders and depth cameras have become increasingly popular in autonomous robotic applications. A common task is to locate and track specific objects of interest located somewhere in the point cloud. This often introduces a tedious network of heuristics to build objects from identified primitives or an intractable high dimensional search space. Through a parameterized object model and certain relaxation functions, a likelihood based view of the data can be used to accomplish these goals with increased performance and reliability. Improvements in mathematics and convergence properties have shown that this method can be realized in real time.