RGB-D correction and completion and its application to SLAM in feature-poor planar environments -

Abstract

This thesis focuses on using an RGB-D sensor (Microsoft Kinect) for localization and mapping in an indoor planar environment. Such environments need special treatment since commonly used algorithms such as feature tracking or Iterative Closest Point (ICP) fail due to lack of visual features and 3D variations within the range of the Kinect. The idea is to make use of the in-range depth data, the 2D appearance data provided by the Kinect, and prior knowledge on the environment (planar) to correct, complete, and extend the 3D information beyond the range of the Kinect sensor. The sensor noise is used to robustly and adaptively fit planes through data points and use a properly designed Markov Random Field (MRF) to label pixels that are consistent with the scene, both in 2D and 3D. As such, depth is inferred at pixels with unknown depth values. After depth correction and extension, an adaptive and robust SLAM method is presented. This novel method uses the new depth data to find the transformation that best registers two consecutive frames. Both feature-point and plane matches are used to improve registration. We evaluate our method on two datasets and show its advantages over other RGB-D SLAM methods.