Camera Tracking Using A Dense 3D Model

Abstract: Abstract In this thesis we present a method for tracking a depth sensor and building a 3D model in real-time. The tracking will be done using the current estimated 3D model. We will show that this approach gives more accurate results and is more robust than the well-known KinectFusion approach. It will also be shown how to colourise the 3D model in real-time. In this work we will study different error metrics such as the projective point-to-point metric and projective point-to-plane metric. Also different norms will be evaluated to find out which gives the best result. We will also show how to represent a 3D model by using a so called signed distance function.