A relaxed Gaussian mixture model framework for terrain classification based on distinct range datasets

The classification of the range (3D) dataset in a meaningful way for the safe navigation of a mobile robot in uneven terrain is still a critical problem. In this paper, we address the problem of classifying a given 3D dataset into a set of known semantic classes, i.e., objects or ground. The classification of range data is carried out using normalized histogram features of 3D information by employing a variational framework of relaxed Gaussian mixture model (GMM). The main merits of the proposed method are independent of prior knowledge about the terrain, range data format and resolution while preserving objects and surface details. The experimental results show that the robustness of the proposed method by achieving classification accuracy of nearly 99.98% on distinct datasets.