Camera Zoom Motion Detection in the Compressed Domain

Abstract

In this paper we investigate the application of local tetra patterns to the compressed domain camera zoom recognition problem. The primary aim is to separate the zooming frames from the non-zooming (panning, tilting) frames for which the block motion vector orientation information is modeled utilizing a 3× 3 neighborhood and local tetra patterns. These patterns are binarized followed by feature reduction using the concept of uniform patterns and finally fed to the C-SVM classifier for recognition purposes. Comparative analysis with state-ofthe-art methods using ESME and H.264 obtained block motion vectors extracted from standard video sequences show superior performance for the proposed method.