E-health and e-monitoring have become an increasingly important area during recent years, being the recognition of motion, postures and physical exercises one of the main topics. In this kind of problem is common to work with a huge training data set in a multidimensional space, so feature selection is absolutely necessary. Most works are based on knowledge extraction using features which permit to make decisions about the activity realized, being feature selection the most critical stage. Conventional feature selection procedures based on wrapper methods or `branch and bound' are highly computationally expensive. In this work, we propose an alternative filter method using a feature-set ranking via a couple of two statistical criteria, which achieves remarkable accuracy rates in the classification process. We demonstrate the usefulness of our method on both laboratory and seminaturalistic activity ambient living datasets for real problems.