This thesis deals with the general question of geometric inference. Given an object that is only known through finite sampling, what conditions are required on the sampling in order to be able to estimate correctly some of its topological or geometric properties ? Topological estimation is by now quite well understood. Most existing approaches rely on the notion of distance function. We use the distance function in order to estimate a notion of curvature due to Federer, that is defined for a rather general class of non-smooth objects. We study the stability of an approximate version of these measures when the unknown object is replaced by a discrete approximation; we also deal with the practical computation of these measures in the discrete...