In this paper, we present a complete framework to produce motion-based geometric segmentation of mesh sequences: each segment corresponds to a mesh region that tends either to move rigidly, or to be stretched in a uniform way. We take a sequence of meshes with varying number of vertices as input and match them by pairs. This allows us to find the approximated displacement vector of each vertex of the first frame all along the sequence. Using these displacement vectors, an initial segmentation is performed on the first pair of frames, clustering vertices into static, stretched and rigidly moving regions. This segmentation is then refined after each matching. Our segmentation is computed on the fly, and lets us find the exact frame when each transformation (rotation, translation, stretch) appears. We have validated our method both on dynamic meshes and unconstrained mesh sequences.