A new method using moments correlation for action change detection in videos

Automated characterization of human actions plays an important role in video indexing and retrieval for many applications. Action change detection is considered among the most necessary element to ensure a good video description. However, it is quite challenging to achieve detection without prior knowledge or training. Usually humans are practicing different actions in the same video and their silhouettes give significant information for characterizing human poses in each video frame. We have developed an approach based on pose descriptors of these silhouettes, cross correlations matrices and Kullback-Leibler distance to detect action changes. In this paper, we will focus firstly on the specific problem of change detection in videos. After that, the proposed approach for action change detection will be detailed and tested on Weizman dataset. Finally, experimental results has been analyzed and showed the good performance of our approach.