Pascal Enderli

Master Thesis
Supervisors: Dr. Danda Pani Paudel, Prof. Angela Yao, Prof. Luc Van Gool

Absolute Visual Localization Using SLAM and a Partial 3D Map

In this thesis, we aim to design a method for an accurate localization of vehicles for dynamic parking lot detection. This is carried out by fusing SLAM and perspective-n-point (PnP) methods within the proposed optimization framework. On the one hand, the 3D map required for PnP is obtained using the Structure-from-motion pipeline, which is then registered to the absolute map by maximizing the consensus set of the GPS-tags of cameras. On the other hand, SLAM is performed using the ORB-SLAM framework. Using the developed method, we are able to correct many failure cases of SLAM-based localization. Additionally, our method also provides the absolute location of the camera at all time.