Exploiting Depth Information from Tracked Feature Points in Dense Reconstruction for Monocular Cameras

摘要

In dense 3D reconstruction work for monocular simultaneous localization and mapping (SLAM), to present coherence and prevent abrupt change in reconstructed surfaces, we normally model the contextual constraint of physical properties in a neighbourhood of space as a certain prior smoothness term concisely into the optimization process. In our work, we first had a careful discussion about the trade-off between precision and accuracy for different prior smoothness terms and how these affected the optimization process of the depth map based on photo consistency measurement. We then presented a method which uses depth information of tracked feature points as priors in the optimization process. Finally, we verified effectiveness of our method by conducting quantitative evaluation experiments in a simulated environment. We also qualitative evaluation in a real environment. We confirmed that feature prior information can improve the accuracy of reconstructed structure at the strong texture area.