Dense 3D Mapping for Indoor Environment Based on Kinect-Style Depth Cameras.

Kinect style depth cameras provide RGB images along with pre-pixel depth information, the richness of their data and recent development of low-cost sensors have made them more popular in mobile robotics research. In this paper, we present a framework of dense 3D mapping. Sparse visual features are used to determine an initial rough transformation, then it is refined by color-GICP (General iterative closest point). We employ a window sparse bundle adjustment to optimize the local map after it is constructed and a new keyframe is created at the same time. Visual features and dense information are also used in loop closure detection, following by a globally consistent optimization based on graph. Moreover, we introduce a user interaction to improve the map building progress. This proposed approach is evaluated by the RGB-D benchmark and two real indoor environments, and experiment results show the feasibility and effectiveness of this approach.