Kinect depth maps preprocessing based on RGB-D data clustering and bilateral filtering

As depth maps obtained by Kinect are noisy and there are invalid areas where are depth values missing in the maps, it´s necessary to discover a proper approach for preprocessing depth maps. This paper proposes a novel method which can process depth maps before using RGB-D data provided by Kinect. In this paper, we abstract the invalid areas in depth map and estimate the size and appearance of each abstracted area firstly. Then, we cluster the RGB image pixels corresponding to the invalid areas of big size and their neighbor areas with MeanShift method. Therefore, the valid depth values of neighbor pixels can be used to fill the missing parts in the same cluster. Finally, this method fills other depth missing pixels and filters the depth map with improved bilateral filtering algorithm. Experimental results show that with the proposed method we can preprocess the depth data with high accuracy and low noise.