Planar modeling of indoor scenes using RGBD images

Abstract

3D registration of scenes is a widely researched topic. Since the availability of low cost sensors like the Kinect RGB-D cameras, they are being extensively used for additional depth information about the scene. Most of the approaches for registration either use point clouds or RGB features to estimate the relative pose between the consecutive views. Our work focuses on using the planar patches of the scenes as primitives for indoor scene registration and modeling. Indoor scenes are structured and can be easily represented using planes. The advantage of using planes is the reduced space complexity from tens of thousands of points to much fewer planes which are present in the scene. Currently ICP is one of the most popular algorithm for scene registration. It uses 3D point clouds for transformation estimation. But the major drawback of the algorithm is that it uses dense points for estimation iteratively which is computationally expensive. Also, ICP requires an initial transform estimate which is used in subsequent iterations which makes it dependent on initial parameters. The proposed algorithm aims to address these issues by using planes for estimation of optimal rotation and translation. General bundle adjustment algorithms use 3D reprojection error between points as the cost function which makes it an expensive step in registration process. We again make use of the planar primitives to reduce the complexity of this step. Finally a 3D model of an indoor scene is obtained from the RGBD images.