Indoor Location Retrieval using Shape Matching of KinectFusion Scans to Large-Scale Indoor Point Clouds.

In this paper we show that indoor location retrieval can be posed as a part-in-whole matching problem of KinectFusion (KinFu) query scans in large-scale target indoor point clouds. We tackle the problem with a local shape feature-based 3D Object Retrieval (3DOR) system. We specifically show that the KinFu queries suffer from artifacts stemming from the non-linear depth distortion and noise characteristics of Kinect-like sensors that are accentuated by the relative largeness of the queries. We furthermore show that proper calibration of the Kinect sensor using the CLAMS technique (Calibrating, Localizing, and Mapping, Simultaneously) proposed by Teichman et al. effectively reduces the artifacts in the generated KinFu scan and leads to a substantial retrieval performance boost. Throughout the paper we use queries and target point clouds obtained at the world's largest technical museum. The target point clouds cover floor spaces of up to 3500m 2 . We achieve an average localization accuracy of 6cm although the KinFu query scans make up only a tiny fraction of the target point clouds.