Continuous and automatic registration of live RGBD video streams with partial overlapping views

Abstract

This paper presents a novel method for automatic registration of video streams originated from two depth-sensing cameras. The system consists of a sender and receiver, in which the sender obtains the streams from two RGBD sensors placed arbitrarily around a room and produces a unified scene as a registered point cloud. A conventional method to support a multi-depth sensor system is through calibration. However, calibration methods are time consuming and require the use of external markers prior to streaming. If the cameras are moved, calibration has to be repeated. The motivation of this work is to facilitate the use of RGBD sensors for non-expert users, so that cameras need not to be calibrated, and if cameras are moved, the system will automatically recover the alignment of the video streams. DeReEs [Seifi et al. 2014], a new registration algorithm, is used, since it is fast and successful in registering scenes with small overlapping sections.