Affordable Access

Adaptives und hybrides SLAM für handgeführte RGBD-Kameras

  • Luu, Thu Huong
Publication Date
Jan 01, 2016
External links


A rising popularity of RGBD sensors caused an increase of research in recording and reconstruction of three-dimensional scenes with such sensors. For reconstruction the so called Simultaneous Localization and Mapping (SLAM) problem needs to be solved. Most RGBD SLAM systems use the point-based Iterative Closest Point (ICP) algorithm. Although the algorithm is well-studied it has problems with noisy data and low-texture areas with few geometric variations, e.g. large and empty surfaces. To address this limitation one option is to additionally exploit planes from surroundings, especially since planar geometry is the most common shape in man-made indoor and outdoor scenes. Taguchi et al. [TJRF13] published the first global registration method combining point-to-point and plane-to-plane correspondences in a real-time SLAM system in 2013. Shortly afterwards Ataer et al. [ACTRG13] extended their approach including a motion prediction model to detect correspondences. A major disadvantage of these methods is the high processing time for a registration step. Thus, the methods are not capable to perform interactive reconstructions. The objective of this thesis is to implement a SLAM algorithm for a hand-held RGBD camera that uses points and planes for registration. In contrast to existing hybrid approaches this thesis pursues the idea of an local registration algorithm. As the number of plane features is significantly smaller than the number of point features in 3D scenes, the algorithm prefers a higher number of planes in calculations. This enables faster correspondence search and registration. A minimal set of correspondences suffices to estimate the sensor pose with the underlying RANSAC based algorithm. This enables the SLAM-algorithm to register featureless regions and environments with few geometric variations where ICP-based approaches would fail. Furthermore the local registration approach enables an interactive use of the system to provide real-time feedback to the user. The registration process is supported by implemented extensions using the detected plane segments to correct geometry. Experiments demonstrate interactive reconstruction of indoor scenes in real-time using a hand-held Kinect. Registration is six times faster than comparable hybrid systems. In addition an improvement over point-based algorithms in textureless regions could be proven using a benchmark for RGBD cameras.

Report this publication


Seen <100 times