Abstract. This paper presents a motion attention model that aims at analyzing gesture saliency using context-related information at three different levels. At the first level, motion features are compared in the spatial context of the current video frame; at the intermediate level, salient behavior is analyzed on a short temporal context; at the third level, computation of saliency is extended to longer time windows. An attention/saliency index is computed at the three levels based on an information theory approach. This model can be considered as a preliminary step towards context-aware expressive gesture analysis.