Classification of Outdoor 3D Lidar Data Based on Unsupervised Gaussian Mixture Models

3D point clouds acquired with lidars are an important source of data for the classification of outdoor environments by autonomous terrestrial robots. We propose here a two-layer classification system. The first layer consists of a Gaussian mixture model, issued from unsupervised training, which defines a large set of data-oriented classes. The second layer consists of a supervised, manual grouping of the unsupervised classes into a smaller set of task-oriented classes. Because it uses unsupervised learning at its core, the system does not require any manual labelling of datasets. We evaluate the system on two datasets of different nature, and the results show its capacity to adapt to different data while providing classes which are exploitable in a target task.