Color-based 3D point cloud reduction

Significant advances have recently been obtained in 3D sensor domain, such as 3D lasers and the Microsoft Kinect, yielding synchronized depth and color information. Nowadays, these sensors are able to collect very dense point clouds. However the interest of such an amount of information is not always justified. The point cloud sampling procedure is an important processing step. It affects the relevance and the accuracy of the remaining steps of this processing chain. In this paper, we present a novel RGB-D down-sampling method. The proposed approach is based on the use of both color information and geometry of the scene. First, a voxelization is performed to maintain the topological details of the scene, then for each voxel, a color based classification of its points is done. Thereafter, only one point of each color class is maintained and all the remaining points are removed. We evaluate the method on the RGB-D data taken from 3D laser scanner and compare it to the method implemented in PCL. The results show that the new method gives better results than the state-of-the-art method. In addition, it opens interesting perspectives to merge color and geometric information.