A depth-incorporated 2D descriptor for robust and efficient 3D environment reconstruction

We develop a system to reconstruct the indoor 3D environment from a sequence of range images. The range images from different partial views of the environment are taken from a range camera while it is moving. The efficiency and accuracy of the 3D reconstruction heavily rely on the partial matching between two range images. We propose a depth-incorporated SURF keypoints and descriptors to stitch two adjacent range image frames. Finally, the stitched reconstruction result is globally optimized through a graph optimization. This effectively reduce the inevitable accumulated error introduced during the pairwise stitching procedure. F5