Recent efforts on geospatial data processing showed a good potential for machine learning, particularly deep learning, tools in the automatic semantic interpretation of a 3D scene. Machine understanding of the surrounding environment is of great importance for several applications, including in particular the development of effective autonomous vehicles. Focusing on the specific case of autonomous driving vehicles, this work considers the problem of automatic segmentation and classification of urban point clouds. To be more specific, this paper considers a mixed approach, where, once properly removed ground points, 3D data segmentation is based on the Euclidean distance, whereas the classification of objects is based on a two-step procedure: first, PointNet++ is used for an initial soft-classification. Then, classification probabilities outputted by PointNet++ are used in combination with some additional geometric features extracted from the point cloud as inputs for a Random Forest classifier. The proposed approach is tested on a dataset collected in Sesto Fiorentino (Italy), showing quite promising performance. Interestingly, the employed tools are available as open-source software.