In this article we conduct an evaluation of feature extraction methods for the problem of human motion detection based on 3-dimensional inertial sensor data. For the purpose of this study, different preprocessing methods are used, and statistical as well as physical features are extracted from the motion signals. At each step, state-of-the-art methods are applied, and the produced results are finally compared in order to evaluate the importance of the applied feature extraction and preprocessing combinations, for the human activity recognition task.