Content-Based Retrieval of Human Actions by Extracting theMain Moving Directions and Their Scales

Beside the fast growth in video media on the Internet, managing these videos is required. To this end, a considerable interest research, namely Content-Based Video Retrieval (CBVR), was introduced to deal with the large amounts of collecting video media. Due to relating a large ratio of these videos to humans, human action retrieval is considered as a new CBVR domain topic. In this paper, we seek to improve the current state-of-the-art CBVR retrieval algorithms accuracy with minor computational cost. In this method, n local features of each video are extracted and each point’s main moving direction and scale is represented by a vector of m dimensions. Then, each video vectors are grouped into four clusters which their centers are considered as the main scales and directions for this video. Then difference of two videos is measured by using their group centers. The experimental results on UCF YouTube dataset with 11 action categories illustrated that in contrast to the Bag-of-Words model our method can perform better.

Content-based video retrieval; Human action recognition; Local point; Vector