Enhancing action recognition in low-resolution videos using dempster-shafer's model

Abstract

With the motivation of lower recognition performance as the resolution of processed action videos decreases, this paper presents a robust action recognition approach based on Dempster-Shafer (DS) theory with assumption that single video frames are independent for action discrimination. By the use of artificial neural network (ANN) estimators trained using single video frames, we first compute the basic belief assignment (BBA) for each video frame in the given query video. The Dempster's rule is then used to combine the resulting BBAs for a final threshold-based decision making. Through experiments conducted on extensive testing data with various levels of video resolutions, we demonstrated outperforming recognition performances by the proposed framework compared with state-of-the-art classifications using sequence matching, voting-based strategy and bag-of-words (BoW) method.