Human actions retrieval from video databases according to the temporal feature by using multiple SVM and SIFT descriptor

Recently, as the amount of videos are raised, video retrieval is going to be a challenge. Videos' contents are very different, so this paper focuses on human actions retrieval, which is useful in different systems such as video surveillance. In this paper two methods for video retrieval process is presented and compared. The SIFT descriptor, used as feature and video's key frames, are selected according to this feature, in the proposed method. The feature vector of a key frame and its neighbors are gathered to form a video temporal descriptor. This descriptor used as a similarity measure in the retrieval process. The second proposed method, uses SVMs to classify the entire dataset and to assign a proper class to the query video. The results of these two methods are compared with the other methods. The Precision and Recall of the proposed method increased by 10.5 and 35.3 percent respectively, as a result of using the video temporal descriptor and SVM classification.