Compressed domain zoom motion classification using local tetra patterns

Abstract

In this paper, a novel compressed domain method for classifying zooming motion is presented. Camera zoom motion classification is an important problem in video analysis wherein the task is to recognize and separate zooming-in camera from zooming-out camera. In our study, we address this problem utilizing local tetra patterns which has earlier found applications in image texture analysis and content-based image retrieval. Towards this goal we model the motion vector orientation and magnitude using local tetra patterns followed by histogram formation. Since the feature dimension is large, uniform pattern-based feature reduction is applied on the histograms to form the feature vector which is fed to the C-SVM classifier for training/testing purposes. Experimental testing utilizing standard video sequences with block motion vectors coming from exhaustive search motion estimation algorithm as well as H.264 obtained block motion vectors along with comparative analysis carried out with existing techniques shows superior performance for the proposed method.