Hybrid 3D registration approach using RGB and depth images

We propose a novel technique for registration of 3D point sets using both the RGB data as well as the depth data. The main advantage of any RGB-D sensor is the pixel wise correspondence between RGB values and depth values, which can be leveraged to register two RGB-D datasets. RGB images are used for correspondence identification and these correspondences are transferred to depth images to be used for the registration algorithm. RANSAC is used for rejection of noisy data points, which increases the registration accuracy. We also analyze and present an error threshold selection strategy for fitting 3D points. Our approach achieves faster execution, thus enabling real-time implementation of change detection and 3D mapping of the environment, etc. Multiple feature extraction methods have been tested to evaluate tradeoffs between accuracy and time.