Preproessing and modeling for visual-based 3D indoor scene reconstruction

In this paper, we present a visual-based method for 3D modeling of indoor scene via mobile robot equipped with a depth camera, which can both acquire color image and dense point cloud. The raw data obtained from depth camera are noisy and non-uniformed, so a set of preprocessing methods, which consists down-sampling with volumetric pixel grid filter, statistical-based outlier removal, moving least square-based interpolation, is conducted to enhance and consolidate the data. A combination of scale invariant feature transform (SIFT) features and iterative closet point (ICP) are performed to estimate the pose of robot for frame alignment. The matched SIFT feature pairs calculated on color images of two frames are used to compute a rigid transformation matrix, which is considered as the initial transformation matrix estimation of ICP algorithm. A dense map is built by aligning of multiple frames, and a compact surface model is achieved by surface reconstruction on dense map using greedy triangulation method. Experiment result showed the method is easy to apply for indoor scene reconstruction and can be executed in nearly real time.