Trilateral filtering of range images using normal inner products

Accuracy of Kinect like range images is limited due to the infrared laser projector mode, so, in mobile robot applications, it is needed to preprocess them to achieve high quality range images without missing data. In this paper, we employ the cosine of the angle between neighboring normal vectors to determine edges and desired regions of range image that the direction of normal vector changed in comparison with neighbors pixels. We propose a novel Trilateral filter which uses the edge regions, detected in the previous step, as direction similarity and distance from plane fitted across the normal vector of central pixel for neighbor´s pixels as distance similarity Gaussian filter beside spatial and intensity similarity Gaussian. By using this filter we can preserve perceptually important `sharp edges´ or `boundaries´ in a range image while reducing the noise or `small details´. Experimental results on several simulated 3D data show that proposed method gains 0.0102 meter in RMS and 87.2151 dB in PSNR criteria.