Improving Dense Real-Time 3D SLAM Using Sparse Geometric Constraints

This thesis explores the extension of a state of the art dense RGBD SLAM system to include detected geometries as elements in the estimated global map. The existing approach leverages an algorithm for dense visual odometry, which is analyzed in detail and reimplemented. It is demonstrated how the inclusion of these detected geometries can improve the state estimate and reduce reconstruction error. These geometric map elements contain invaluable semantic information about scene content that more dense map representations lack, and serve to improve localization, reduce dense reconstruction error, improve scene understanding.