Abstarct: Nowadays with the increasing concern on security problems and necessity for studying human activitiy, requests for an intelligent system which automatically recognizes human activities increases. In manual inspection in which a person watches given video for recognizing human activities is inefficient mainly because of tiredness. Hence the need for an automatic system for activity recognition is demanded more than before. Activity recognition is a challenging task due to camera motion, clutter in video background and inter personal differences. Another challenge in human activity recognition is performing an activity in different forms. Thus, we need utilize semantic information about the activity that occurred and address these challenges. Unlike low level features, semantic features describe inherent characteristics of activities. Therefore semantics make the recognition task more reliable especially when the same activities look visually different. In this thesis, we propose an approach for extract semantic features and utilize them aligned by non-semantic features for human activity recognition. In the proposed method dense trajectories, HOG, HOF and MBH are used as non-semantic features. Then salient trajectories are determined. Salient trajectories are clustered using two laxyers clustering method then the obtained clusters construct semantic features. These features are encoded by bag of word algorithm. Finally by using the KNN classifier activities are recognized. The implementation of the proposed method result in 92.9% average accuracy on UCF sports dataset.