This paper describes preliminary results of a research performed in the framework of the Enactive project (EU IST NoE Enactive). The aim of this research is to improve believability of a virtual character using qualitative analysis of gesture. Using techniques developed for human gesture analysis, we show it is possible to extract high-level motion features from reconstructed motion and to compare them with the same features extracted from the corresponding real motions. Moreover this method allows us to evaluate whether the virtual character conveys the same high level expressive content as the real motion does, and makes it possible to compare different rendering techniques in order to assess which one better maintains such information.