In this article, we propose a new approach for 3D object classification, based on the Time Series Shapelets of Ye et Keogh (2009). The main idea is to use discriminants sub-parts for the supervised classification in order to take care of the local nature of pertinent elements. This allows the user to be aware of these sub-parts which have been useful to determine the corresponding class of the object. Final results confirm the interest of random feature selection for pre-selection of attributes in supervised classification.