Despite their attractive properties of invariance, robustness and reliability, fuzzy directional features are not hitherto paid the attention they deserve in the activity recognition literature. In this paper, we propose to adopt an innovative approach for activity recognition in real-world scenes, where a new fuzzy motion descriptor is developed to model activities as time series of fuzzy directional features. A set of one-vs.-all SVM classifiers is trained on these features for activity classification. When evaluated on our dataset (i.e., IESK action dataset) incorporating a large and diverse collection of realistic video data, the proposed approach yields encouraging results that compare very favorably with those reported in the literature, while maintaining real-time performance.