Range image denoising using a custom built multiscale transform

Range images are valuable sensory data which play an important role in spatial awareness of a mobile robot. Advent of cheap range cameras which were originally intended for gaming, vastly increased the access to such 3D data; although the quality is still poor for many robotic tasks. Considering the dominant planar geometry of indoor environments, in this paper a custom transform based on lifting scheme is proposed to enhance range readings and enforce potentially present planar model. Our custom build transformation decomposes input range data into approximation part which represents planarity consistent data and details which represents non-planar data. Thanks to such an effective representation, we discriminate real edges from noise data by a simple thresholding method. Detail coefficients are compared to threshold value and nominal details are suppressed. Enhanced range data are reconstructed using approximation part and remained details. Several experiments demonstrate discriminative potential of our proposed algorithm. Finally, remaining challenges and future works are described.