In the last decade, Content-Based Video Retrieval (CBVR) was introduced to handle the large amount of collected videos on the Internet. Due to relating a large amount of the videos on the Internet to humans, human action retrieval is supposed as a new topic in CBVR domain. In this paper, we seek to improve the current state-of-the-art retrieval algorithms for CBVR by utilizing the fuzzy clustering with less computational cost. In this method, n local points of each video are extracted and each point’s main moving direction and its scale is represented by a vector of m dimensions. Then, by analyzing these local points’ vector for each video, it is clear that deviation between the vectors is less and these vectors are similar. Moreover, by clustering these vectors, the cluster centers can indicate the main moving directions and scales in the video. Due to the less deviation between the vectors of a video, fuzzy clustering can result better cluster centers. The experimental results on UCFYT sport dataset illustrated that, in most cases, our method outperforms the final results of the Bag-of-Words model. Moreover, utilizing the fuzzy clustering can outperforms the results of k-means method.
