This paper presents a data-driven approach to estimate the kinematic performance of tennis players, using kernels to extract a dynamic model of each player from motion capture (MoCap) data. Thus, a metric is introduced in the Reproducing Kernel Hilbert Space in order to compare the similarity between models so that the built kernel enhances groups separability: the baseline reference group and the group including players developing their skills. Validation is carried out on a specially constructed database that contains two main testing actions: serve and forehand strokes (carried out on a tennis court). Besides, the classical kinematic analysis is used to compare our kernel-based approach. Results show that our approach allows better representing the performance for each player regarding the ideal group