Identification results dramatically depend on the excitation properties of the motion used to sample the identification model. Strategies to define persistent exciting trajectories have been developed for manipulator robots with few DOF. However they can not easily be extended to humanoid systems and humans due to the important number of DOF; and empirical knowledge is often used to generate and select persistent exciting motions. In this paper we propose a method to choose persistent exciting motions from an existing dataset in order to optimize both the identification results and the computation time. This method is based on the use of the identification model of legged systems obtained from the base-link equations. Instead of using well-established consideration on the condition number of the regressor matrix, the method uses a decomposition of the regressor into elementary sub-regressors and the computation of the condition number for each. A selection rule is then proposed. The overall method is experimentally tested to identify the human body inertial parameters using a data-set of 40 motions. Comparative results obtained from different combinations of motions are given.