A Kinect-Based Front-End for Graph-SLAM Using Plane Matching in Planar Indoor Environments.

We present a pose graph optimization approach which enables a mobile robot to create 3D maps of planar indoor environments using a Microsoft Kinect camera. Rather than using point features, the approach relies on the registration of planar surfaces. Plane matching is used to implement a (front-end) for the construction of a pose graph, which is then optimized by a state-of-the-art (back-end). Vertical planes are extracted from acquired point clouds associated to the poses of the robot; then a plane matching algorithm is used to create constraints among successive robot poses. Place revisiting episodes are detected using 3D features in order to provide loop-closing constraints. These constraints provide the input for a pose graph optimization algorithm, which computes an estimate of the robot trajectory. Finally, the 3D map is created by attaching to each pose of the trajectory the corresponding planes. Planar surfaces are more robust and descriptive with respect to point features and provide an accurate estimate of rotations. Moreover, the front-end combines geometric and appearance-based information to filter out outliers and perform robust plane association. Preliminary experimental results in real environments show that the approach is able to create 3D maps which are consistent and close to reality.